Summary: User researchers as role-playing game classes | AI posture recommendation feature in Huawei Pura 90 phones | AI helps teaching assistants provide better feedback to students | AI as a metacognitive coach | Recommending an entertaining short AI film | SpaceX’s plan for 100 TW orbital compute | New Chinese video model Happy Horse from Alibaba

UX Roundup for May 4, 2026 (GPT-Images-2)

User Researchers as Role-Playing Game Classes

I continue to play with the new capabilities unlocked by the GPT-Images-2 model’s reasoning abilities and world knowledge. Here are examples of the prompt “draw the class sheet for ‘user researcher’ as a new RPG class.” I specified different subclasses for these class sheets.

The manager. (GPT-Images-2)

Discount usability. (GPT-Images-2)

The insight oracle. (GPT-Images-2)

Diary studies. (GPT-Images-2)

The field ethnographer. (GPT-Images-2)

ResearchOps. (GPT-Images-2)

I made these class sheets for fun, but they can serve a serious purpose for teams that need to introduce themselves and their specialties to stakeholders or colleagues in other disciplines. You can upload your own photo as a reference image and ask for the player to be drawn with your likeness.

I built these across several chat sessions, because it was too much fun to stop, which is why the styles differ across cards. If you build everything within a single session, or if you upload existing cards as a style reference, GPT delivers strong design consistency. (On the other hand, within a session, GPT drew characters that looked too similar, overdoing style consistency.)

AI Posture Recommendation Feature in Huawei Pura 90 Phones

I have often talked about the need for AI features to meet users where they are and to smoothly address a current need, rather than being bolted on for the sake of saying “now AI-enabled.”

The new Huawei Pura 90 phone series provides a good example. Unfortunately, Huawei remains banned in the USA for political reasons, but luckily, we can see the feature demonstrated on X now that it is a fully international social media platform with automated translations. (Which serves as another example of integrated AI.)

Watch this short video posted by Indian user “Noah Cat” of Huawei’s “AI Posture Recommendations” in action. It looks like a promotional video originally produced by Huawei and probably posted to some of the Chinese social media platforms.

Still image from the Huawei demo video. The text by the arm holding the selfie-camera reads “raise your arm” and the text to the right reads “keep smiling,” the latter probably being redundant advice for a selfie. The part of this UI that I really like is the way the outline naturally helps you position yourself in the photo to achieve an engaging composition.

AI Posture Recommendations overlays real-time pose outlines in the camera viewfinder, suggesting scene-tailored positions like raised arms or dynamic stances to improve photo quality on the first try. Sorry to be prejudiced but this does seem like a feature targeted at young people in Asia who really love funny postures for their selfies. (As empirically observed every time I’m in Asia.) However, serving your target audience is good UX.

For my purposes, the question is not whether funny poses in selfies are good or bad. My point is that using AI to analyze how a user could look funnier and then overlaying that posture recommendation directly where he or she is already looking is an AI UI with great usability.

This is a command-free interaction that’s integrated with the task and appears exactly when needed.

No commands: just AI that appears naturally integrated with the task when it’s needed. And fun too, which is what we need to improve public perceptions of AI. (Nano Banana 2)

AI Helps Teaching Assistants Provide Better Feedback to Students

Many studies have already shown that students’ learning suffers when they use AI to do exercises instead of solving the problems on their own. (AI does help students when used as a coach or teacher, to help the students get over stumbling blocks when they are solving problems.)

But what about doing some of the teachers’ work on their behalf? In one study, this did help the students, by providing better feedback on their class assignments.

A research team at the University of Michigan tested an AI system called FeedbackWriter that augmented teaching assistants’ (TAs) feedback on assignments college students. Their randomized controlled trial spanned a computer‑science course with 354 students and 12 teaching assistants. Half of the TAs used the FeedbackWriter tool, which generated draft comments and allowed TAs to edit and approve them; the control group wrote comments manually.

The results were striking. Students whose assignments were graded with AI‑drafted feedback made more substantive revisions and improved the clarity, correctness, and completeness of their essays compared with peers receiving traditional feedback. TAs using the AI found that it encouraged them to provide more balanced feedback, highlighting strengths as well as weaknesses. Importantly, the AI never operated autonomously; human oversight remained central, preventing hallucinations and ensuring alignment with grading rubrics. In post‑study interviews, TAs emphasized that the AI acted as a brainstorming partner. It suggested phrasing for difficult comments and helped them notice patterns across student mistakes. The researchers argue that by embedding AI at the right point in the workflow to support but not supplant human judgment, FeedbackWriter increased both the quality of feedback and instructor satisfaction.

We know that one thing AI is perfect at is to generate word count. It’s no great surprise that the feedback to the students was lengthier when the TA used AI: 284 words vs. 159 words. The AI-augmented feedback also broader, covering 20% of scoring rubrics on average for the AI-augmented feedback and only 16% in the no-AI condition.

Of course, the extra feedback from AI vs. humans-only might have been misleading or overwhelming, but this was not the case, according to the scoring of the final assignments handed in by the students after they had a change to use the TX feedback to improve their drafts.

The quality of the students’ final draft was rated as 29% higher when their TA used AI to generate feedback on their first draft. (The rating between the two drafts increased by 22 points in the AI-augmented condition and only 17 points in the no-AI condition.)

The study was at the University of Michigan, not a medical castle, but this is roughly how it played out. (Nano Banana 2)

AI as a Metacognitive Coach: Curing Cognitive Laziness

There is a pervasive fear in the tech community that artificial intelligence is turning us (and the next generation of students) into lazy thinkers. If Codex can instantly spit out a fully functional Python script or Claude Designer make a high-fidelity UI prototype, won’t we lose our fundamental problem-solving skills? If AI acts as an over-eager intern doing all the grunt work while we merely nod and approve, cognitive atrophy seems inevitable.

But a new randomized controlled trial published in Frontiers in Psychology (April 2026) by Hou et al. from the Luoyang Institute of Science and Technology in China flips this pessimistic assumption on its head. The researchers discovered that when AI is designed correctly, it doesn’t diminish human agency. It actually cures cognitive laziness.

The researchers studied 122 computer science undergraduates learning to program. As UX professionals know, there is a vast gulf between what users say and what they do. In surveys, students consistently claim they know how to plan, check their work, and debug systematically. But faced with a blank screen, they panic and start mashing the keyboard in a frenzy of trial-and-error.

To bridge this awareness–behavior gap, the researchers used a customized programming environment with an integrated autonomous AI agent. The unexpected, brilliant twist? The AI was expressly forbidden from writing a single line of code.

Instead, it acted as a behavioral scaffold. It quietly monitored students’ fine-grained interaction logs for signs of impulsive behavior. If a student opened a complex assignment and instantly started typing, the AI intervened with a lightweight popup: “You started coding very quickly. Have you broken down the problem requirements?” If a student triggered three errors in a row and rapidly changed the code without thinking, the AI would prompt: “Repeated errors detected. Instead of guessing, consider reading the traceback message.”

Slower is Better: The Power of Friction

The results run contrary to the traditional usability dogma that “faster is better.” In this case, good UX meant introducing deliberate cognitive friction. The AI successfully inhibited the novice’s “rush-to-code” impulse. The time students spent planning before their first keystroke quadrupled from 45 seconds to 186 seconds.

Because they paused to plan, their blind guessing plummeted from 36% to just 12%, and students’ debugging success rate soared. They engaged in deeper cognitive struggle than the unassisted students, ultimately achieving significantly higher academic scores. They didn’t thrash; they thought.

The reason slower was better in this study is that the goal of education is to learn, not to get the assignments done as fast as possible.

Findings: How Students’ Behavior Changed

1. Objective Behavioral Shifts (The Log Data): The system didn’t change the quantity of the students’ actions: both groups executed their code roughly the same number of times and had similar total debugging counts. But the AI radically improved the quality of those actions.

• Enforced Planning: As noted earlier, unassisted students rushed into the task, typing their first line of code in just 45 seconds. The AI-coached group was nudged to slow down, taking an average of 186 seconds before their first edit. They actually stopped to design an architecture, reflected by their comment-to-code ratio jumping from a dismal 3.5% to a robust 15.2%.

• Deep Error Diagnosis: When encountering a runtime error, the control group paused for an average of just 6.2 seconds before changing their code, a classic hallmark of frantic novice “thrashing.” The AI group averaged an 18.5-second “error dwell time,” utilizing the pause to actually read the traceback data.

• Reduced Blind Guessing: Because they stopped to think, the AI group’s “ineffective trial rate” (blindly tweaking code without understanding the problem) plummeted to 12.4%, compared to an alarming 35.8% in the control group. Consequently, their debugging location accuracy leaped from 61.6% to 72.4%.

2. Superior Academic Outcomes: This deliberate behavioral friction did not hamper their actual learning; it accelerated it. In an independent post-test programming assessment where the AI scaffold was removed to test true learning transfer, the experimental group scored an average of 85.2 out of 100, decisively beating the control group’s 76.5. The researchers found a strong correlation confirming that the deliberate debugging behaviors fostered by the AI directly translated into higher independent academic achievement.

3. Surging Metacognitive Awareness: The intervention also fundamentally altered how students perceived their own cognitive abilities. On a 100-point Metacognitive Awareness Inventory scale, the two groups started with nearly identical baseline scores. However, by the end of the study, the control group had gained only 10.9 points, while the AI-assisted group surged by a massive 33.4 points.

The Superiority of Teaching Metacognitive Behaviors

What the AI taught these students wasn’t Python syntax. It taught metacognitive behaviors: the ability to plan, monitor, and regulate one’s own thought processes.

Traditional education still focuses on drilling plain facts and rote skills into students. But by the time today’s university freshmen graduate, AI advances will have rendered their hard-earned skills entirely obsolete. We are rapidly entering the third user-interface paradigm in computing history: intent-based outcome specification. If you can specify the outcome, the AI will execute the skill.

What will not be obsolete is metacognition. Teaching a student the exact syntax of a “for” loop is a fool’s errand; developing the metacognitive discipline to pause, plan, and verify is how you forge a professional who can effectively direct future AI systems.

AI as a Coach

This study perfectly validates my recent article on the 4 Metaphors for Working with AI, where I categorized our interactions into four roles: Intern, Coworker, Teacher, and Coach.

Most educational anxieties stem from the Intern metaphor. When we hand an AI a prompt and it does the work, it completely undermines learning because the AI just solves the problem for the student.

Other systems act as a Teacher, patiently explaining concepts like a personalized tutor. This is the metaphor used with most other uses of AI in education today. Much better! But even this focuses on delivering soon-to-be-obsolete facts.

The educational AI in this study operated strictly under the Coach metaphor. A gymnastics coach doesn’t do the floor routine for you. The coach stands on the sidelines, watches your form, and yells, “Straighten your left leg!” By focusing on the user’s process rather than the product, the AI coach forces the human to do the heavy lifting, but with superior form.

Takeaways for Product Design

For product design strategists, there are two major takeaways here:

1. For your personal career future: Look closely at your daily tasks. If your primary economic value is generating standard deliverables, whether boilerplate code, simple wireframes, or routine usability reports, you are competing with the AI Intern. And the AI Intern works faster and cheaper.

Your future viability relies entirely on elevating your metacognitive skills. You must become the strategist who understands how to plan a multi-phase research initiative, monitor the changing market landscape, and regulate your team’s approach. Cultivate your ability to direct the AI, monitor its outputs critically, and adjust your strategy.

2. For your design projects: The tech industry is currently obsessed with the generative AI “magic button” that does all the work with zero effort. But if you design enterprise software, learning systems, or professional UX tools, consider the AI-as-Coach paradigm. How can your AI monitor user behavior to offer non-directive, process-oriented prompts?

Well-designed friction is a core component of human–AI interaction. Sometimes, the best UX doesn’t come from doing the work for the user, but from providing the strategic scaffolding that empowers them to master the work themselves.

I think Socrates might have liked this new approach to using AI to help students learn how to think and solve problems. (GPT-Images-2)

Entertaining AI Short Film

I recommend watching The adventures of Kulrik and Boon part one - Journey to Ashkarn (YouTube, 9 min.), a short fantasy film by Stevie Mac about the adventures of an orc and his trusty giant pig. Watch in 4K!

The film is entertaining in its own right, with everything from exciting vistas to fun moments where the orc interacts with his pig. However, we should recognize that the story only holds up for a short film and would probably not work if expanded into a two-hour feature film. Such is the state of the art in AI-filmmaking this year. Next year? Mid-length episodes, and by 2028 I expect to see full-length movies made with AI that equal the average output of the legacy studios, if not their very best work.

I am already looking forward to the second episode of Kulrik and Boon. Mostly for the pig. (GPT-Images-2)

For now, shorts are the first native media form of AI.

The short film was made with Seedance 2.0, which is the best current AI video model, and almost every scene looks superb. (In one scene, the orc fights a bunch of skeletons that have come alive to defend a treasure, but fighting or not, I don’t think skeletons spout blood when beheaded.)

The broader lesson is that AI changes the production bottleneck. For a century, moving images required capital: cameras, crews, locations, lighting, actors, editors, and coordination. AI video attacks this bottleneck directly. The scarce resource is no longer equipment. It is ideas and storytelling.

China gives us the best early evidence of what happens when this shift meets a vast short-video culture. AI animated shorts have become a mainstream category there, with Douyin alone reporting 76 billion cumulative views across 30,000 of titles. This is the right incubation environment. Short videos reward novelty, speed, and emotional clarity. They punish overbuilding. A creator does not need a perfect cinematic universe. He or she needs one vivid premise, three good scenes, and a payoff before viewers swipe away.

This is why AI video is important for individual creators and small teams. The technology does not merely lower costs; it lowers organizational complexity. A single person can try ten visual directions in the time a team might spend negotiating one. A two-person team can combine writing, art direction, editing, and marketing without a studio hierarchy. The result will not always be good. Most of it will be dreadful, because easier production increases the volume of weak work. But it will also increase the number of experiments.

This matters. Many great media formats begin as cheap, slightly embarrassing experiments: comic strips, radio serials, YouTube channels, podcasts, TikTok sketches. AI shorts are the next such format. They let outsiders make visible what remained trapped in their heads. When production becomes cheap enough, talent no longer needs permission. It only needs distribution, persistence, and judgment. China is showing the world that the small creator, not the big studio, may be the first native author of AI video.

New media forms never live up to the standards of the old gatekeepers. Short videos are a native media form of the AI era and will get better, even as old-school film reviewers and movie directors denigrate them. (GPT-Images-2)

SpaceX to Build 100 TW Orbital Compute

The world is currently suffering a terrible compute famine, delaying the growth of AI and making advanced AI too expensive for many applications. The short-term solution is to build more data centers and more power plants, using any type of generation that can be constructed fast, whether solar, nuclear, or (sadly for global warming) gas-powered turbines.

However, even the most aggressive plans will be insufficient to power superintelligence at the scale we will need to revolutionize the world economy and give everybody higher living standards, better healthcare, and individualized education that actually brings each student to his or her potential. Just imagine running every school in Africa on AI and giving every student an AI tutor that’s many times more powerful than GPT 5.5 Pro at extended thinking.

100x more compute won’t do the job.

That’s why I am happy that SpaceX is thinking big and launching a plan to build 100 TW of orbital compute by 2045. Currently, the biggest AI supercomputer is that company’s Colossus, which is rated at 2 GW of compute. Thus, the new plan calls for 50,000x more compute than the world’s biggest right now.

The plan to build 100 TW of compute, running on solar power in orbit around the Earth. (GPT-Images’2)

The plan for orbital compute is included in a new stock incentive plan for SpaceX founder Elon Musk. If the company achieves this audacious goal, he’ll be rich. (Well, he’s already rich, but if he can do it, he will deserve to be even richer.)

Will SpaceX succeed? We’ll have to check back in 19 years, but since Elon Musk is notorious for overly optimistic time schedules, there’s a risk they will be late. However, the exact date doesn’t matter that much to me: what’s important is that somebody is starting a plan to build the vastly larger amount of compute we will need to bring humanity to the next level. This goal will indeed take time, whether 19 or 25 years, which is exactly why we should start now.

New Video Model Happy Horse from Alibaba

China continues to lead in AI video models, with Alibaba’s new Happy Horse model scoring high on the leaderboards and getting rave reviews by many AI creators.

Alibaba’s Happy Horse is having a soft launch at the moment, with availability on some of the API-based creator services, like Magnific and Topview. (GPT-Images-2)

In my limited experimentation, I have been impressed with Happy Horse, and I would rank it just below ByteDance’s Seedream 2.0 and a smidgen above Kuaishou’s Kling 3. This means that the best American video model, Google’s Veo 3.1 is currently out of medal contention.

Happy Horse does well in cinematic scene composition, multi-shot sequences, and lip synch and speech generation. However, when I made a test cut with a snooty English butler and explicitly specified a posh British accent, the butler ended up talking in a distinctly American accent. A small fail, and Happy Horse also had slightly awkward timing in cutting between two camera angles while the butler was saying his lines. (However, Kling often gives me terrible speech generation in English, so I do rank Happy Horse higher than Kling, even if Kling probably has better multi-shots. As always in creative AI, the choice of model should depend on your creative vision. Also Kling generates native 4K video as the only one of the leading video models, so if you need the highest resolution, Kling is the place.)

Happy Horse didn’t make my butler sufficiently English. (GPT-Images-2)

As an aside, I’m happy that we’re finally getting memorable names for AI models rather than just obscure abbreviations (looking at you, ChatGPT) with arcane version numbers (also looking at you, ChatGPT, where o3 was more powerful than 4o — though OpenAI seems to have fixed its numbering scheme as of lately).

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,237 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: Standard usability testing uncovers interface flaws, but deprivation studies reveal if a product should exist at all. By intentionally forcing users to abstain from software, we expose hidden dependencies, useless ghost features, and duct-tape workarounds, providing empirical data to ruthlessly declutter bloated interfaces and maximize true utility.

Deprivation studies are a rarely used, but highly strategic, user research methodology. (GPT-Images-2)

[Watch my short explainer video about deprivation studies (YouTube, 6 min.)]

Traditional user studies remain the most efficient way to discover how people operate an interface, but they rarely show whether the interface deserves to exist. Deprivation studies intentionally remove a product, service, or feature for a defined period, exposing dependencies that ordinary observation misses. By watching the workarounds users build when the product is gone, UX teams get closer to revealed behavior than self-report. The result is a sharper basis for prioritizing features, restructuring information architecture, and removing interface clutter.

Human memory is notoriously fallible, and human self-reporting is inherently biased. If you ask a user how they use a piece of software, they will give you an idealized, rationalized narrative. They will describe a logical sequence of events that reflects how they think they should behave, rather than the messy, error-prone, distracted reality of how they actually behave. If you put them in a usability lab and observe them attempting to complete a task, the illusions vanish. You see the hesitation, the misplaced clicks, and the navigation failures. This is why standard usability testing with five users remains the most cost-effective way to find interface flaws. Discount usability engineering works.

You can’t trust what people say about how they use your product. (GPT-Images-2)

But standard usability testing has a blind spot: it studies episodes, not dependencies and habits.

If you want to know whether a specific button is too small, or if a form is too confusing to fill out, a one-hour user test is perfectly adequate. But what if you want to know whether the user actually needs the button at all? What if you need to measure the cumulative, long-term value of a product that a user touches forty times a day?

Standard usability testing studies episodes; deprivation studies reveal the habits underneath them. As an aside, the study facilitator in this cartoon commits a methodology error in telling the user that the task was successful. You need to stay neutral whether the user makes the right or the “wrong” action. (GPT-Images-2)

When a technology becomes truly integrated into a user’s daily life, like a smartphone operating system, an enterprise team messaging app, or a digital calendar, it ceases to be an external tool and becomes an extension of the user’s cognitive architecture. The interface becomes practically invisible to them. The behavior becomes an ingrained reflex.

You cannot study a reflex by asking a user to perform it in a lab environment. You cannot accurately measure the value of a habit by asking the user to fill out a market research survey. Users cannot reliably report on the value of the air they breathe until you start suffocating them. To understand the true utility of a pervasive, habitual technology, you must disrupt the habit entirely.

You take the product away. This is the premise of the deprivation study.

In a deprivation study, participants agree to refrain from using a product or technology for a specified duration. (GPT-Images-2)

Deprivation is not a more sadistic version of usability testing. Its real power is counterfactual. It changes the research question from “Can users operate this interface?” to “What collapses when this interface disappears?” The unit of analysis is no longer the screen. It is the user’s ecosystem of memory, coordination, motivation, social obligation, and workaround labor. A good deprivation study is therefore less a test of a product than a stress test of a way of working or living.

A product or service may appear to serve a simple function, while its real value is the social behavior it enables. (GPT-Images-2)

The Fallacy of Self-Reported Value and Analytics

Many designers and product managers still mistake usage analytics for user value. High usage can reflect value, but it can also reflect poor information architecture (users clicking frantically to navigate a mess), corporate coercion, or a user’s inability to find a better path,

Analytics tell you what is happening on the screen, but they are utterly incapable of telling you why it is happening. Even more importantly, analytics cannot distinguish value from captivity. A forced click, a workaround click, a panic click, and a delighted click look identical in a dashboard. Deprivation separates these cases because the user either reconstructs the missing function at great cost, replaces it without regret, or refuses to reconstruct it at all. Those three reactions correspond to value, substitutability, and waste.

Standard analytics can’t tell the difference between “good clicks” and “bad clicks.” You may be celebrating “engagement” increasing, when all that’s happened is that users are more confused and lost. (GPT-Images-2)

To find the “why,” product teams often turn to focus groups. Focus groups are weak for interaction design because the setting rewards performance rather than observation. Participants influence one another, try to sound thoughtful, and produce socially acceptable explanations for behavior that is often automatic, fragmented, and private. If you ask a room full of people how much they value their email client, they will complain about minor UI grievances, such as the color of the unread indicator or the placement of the search bar, but they will completely fail to articulate the deep, systemic ways the tool supports their daily workflow.

The query effect: when you ask people, they will make up an opinion about anything and make it sound as profound as possible. However, these statements often have nothing to do with users’ real behaviors, because the query responses were constructed after the question was asked. (GPT-Image-2)

Surveys are no better. Present users with twenty proposed features to rank, and they will request a confusing array of bells and whistles they will never actually use, simply because the psychological cost of saying “yes” to a hypothetical feature is zero.

When there is no cost to asking for more features in a survey, users will seem to want feature bloat that would be unhealthy for them in reality. (GPT-Images-2)

Deprivation studies cut through the noise of vanity metrics and self-reporting. By enforcing an artificial scarcity, you force the user to confront their actual, functional reliance on your system. You bypass their rationalizing brain and measure their behavioral reality.

This is why the researcher must distinguish dependency from value. A user may suffer because the product is genuinely useful, but also because the product has captured critical data, imposed a proprietary workflow, trained collaborators around its notifications, or made alternatives unavailable. The right question is not merely “Did users miss it?” The better question is: “What kind of dependency did the product create, and is that dependency desirable?” (And when I say “desirable,” that’s from the perspective of the users and their companies, not from the perspective of vendor lock-in which won’t persist in the age of AI making it cheap for competitors to launch services that don’t hold their customers hostage.)

A Silicon Valley saying is that many SaaS companies don’t have customers, they have hostages, who are kept trapped by locked-up data. But if people detest your service, they will eventually leave as AI creates alternatives. (GPT-Images-2)

What Exactly is a Deprivation Study?

Instead of asking users to use a product and observing their success, a deprivation study asks them to stop using it and observes the resulting failure. By artificially inducing a state of technological absence, researchers can study the gap left behind.

The methodology traces its roots to early media and sociological research. In the late 20th century, researchers paid families to physically unplug their televisions for a month to document the profound shifts in household dynamics, time allocation, and communication. In modern UX research, we apply this same blunt-force methodology to pervasive digital technologies.

Throughout the deprivation period, researchers collect qualitative data on the participant’s experiences, their emotional states, and, most importantly, the alternative methods and workarounds they employ to accomplish their goals. The study operates on a very simple, pragmatic premise: You do not know what a complex system actually does until you watch it fail. By entirely removing the system, you force a failure of the user’s established workflow. Suddenly, every micro-task that was previously handled by the invisible interface must be consciously addressed by the user. The massive friction they experience and the solutions they invent provide high-fidelity data about the product’s true utility.

When to Use Deprivation Studies (And When to Avoid Them)

Deprivation studies are not a universal tool. They are costly, logistically demanding, and intentionally disruptive. Do not use deprivation to answer a question that a one-hour usability test, analytics review, diary study, or customer interview can answer more cleanly.

Use deprivation when you need to know what role a product plays in real life. This is common before a redesign, before a pricing change, before removing a feature, before repositioning a product, or before entering a market where users already have entrenched alternatives. It is also useful when a product team is drowning in feature requests and cannot tell which features are truly essential. The study can reveal whether users miss a feature, replace it easily, or discover that they never needed it.

You should deploy deprivation studies only when your strategic research questions involve deep behavioral integration and product value.

Ideal Scenarios for Deprivation Studies

1. Assessing High-Frequency, Habitual Products: This methodology shines when investigating products that users interact with daily. Think of ubiquitous consumer applications like social media networks, fitness wearables, and streaming services, or essential enterprise tools like Slack, Jira, or Salesforce. The product must be embedded enough in the user’s routine that its removal causes an immediate, measurable disruption. If they do not have a habit, you cannot deprive them of it. Deprivation studies are particularly strong for mature products because mature products become infrastructural. Early in a product’s life, users can still describe the novelty. Later, the product disappears into routine. Deprivation makes the infrastructure visible again.

2. Diagnosing Creeping Featuritis and UI Bloat: Over time, successful software inevitably becomes bloated. Product managers succumb to the temptation to add “just one more feature” to satisfy a vocal minority of enterprise clients or to match a competitor’s marketing checklist. The interface becomes cluttered, cognitive load increases, and overall usability plummets. Deprivation is the ultimate diagnostic tool for featuritis. When you take a specific module of the product away, you discover exactly which features the user actually misses, and which ones they never even realized were gone.

Deprivation reveals which features users would fight to recover and which ones were only decorating the interface. (GPT-Images-2)

3. Evaluating the True Competitive Landscape: You may believe your primary competitor is another software company. Often, your most dangerous competitor is a physical legal pad, an Excel spreadsheet, or a quick phone call. Deprivation studies force users to rely on fallbacks. By observing these fallbacks, you discover the true competitive landscape of the user’s workflow.

4. Prior to High-Stakes Overhauls: If your organization is planning a radical visual and structural overhaul of a core product that risks alienating your existing, loyal user base, a deprivation study serves as a vital reality check. By understanding precisely what specific functions users rely on to survive, you can ensure that the fundamental utility of the product is preserved and protected, even if the navigational paradigms change completely.

When NOT to Use Deprivation Studies

1. Testing Unreleased Prototypes: You cannot deprive a user of a tool they have never used. The methodology relies on breaking an established mental model.

2. Evaluating Low-Frequency Transactional Sites: If a user only interacts with your software once a year, such as annual tax-filing software or a car registration portal, a two-week deprivation study will yield absolutely zero data. The user was not going to use it anyway.

3. Mission-Critical Health, Safety, and Financial Systems: You must never violate the ethical baseline of user research. Do not conduct a deprivation study if taking the product away will cause physical harm, significant financial loss, or severe professional damage. You cannot ask an air traffic controller to participate in a deprivation study of their radar screen. You cannot ask a diabetic user to stop using their continuous glucose monitoring app. The friction you introduce must be highly annoying, but it must not be catastrophic. Common sense must prevail above all research curiosities.

Furthermore, do not use total deprivation when partial deprivation would answer the question. If the team wants to understand the value of reminders, participants do not need to abandon the entire calendar. They may only disable reminders, or use a paper calendar for selected appointments, or suppress notifications during certain hours. Smaller interventions often produce cleaner findings and will reduce study abandonment.

Think of deprivation as a dosage problem. Total abstinence is the maximal dose, and it is often unnecessarily crude. In many mature systems, the better intervention is selective removal: remove autocomplete but not search, notifications but not messages, dashboards but not raw records, mobile access but not desktop access, AI suggestions but not manual authoring. Different removals reveal different forms of value. A missing notification tests interruption dependency. A missing search function tests recoverability. A missing dashboard tests whether summary views genuinely guide decisions or merely decorate the interface.

Selective deprivation is easier to enforce and gives you valuable information about how people depend on that one thing you have removed. (GPT-Images-2)

The Anatomy of Absence: Five Main Findings to Expect

When you run a standard usability test, you find out if a user can successfully click a button. When you run a deprivation study, you find out why they needed to click it in the first place. The findings you extract from this methodology will not look like a standard bug report. They will reveal deep, structural insights about human behavior.

The findings usually fall into five categories. For all five categories separate task pain from memory pain, coordination pain, emotional pain, and identity pain. Many products are misdescribed because teams assume they are selling the task layer. In reality, users may value the product because it remembers for them, synchronizes a group, legitimizes a decision, calms anxiety, or gives them confidence that nothing has fallen through the cracks. Deprivation reveals the layer of value the product team has been too close to see.

1. The “Phantom Limb” Phenomenon and Reflexive Habits

When a digital product is deeply ingrained in a user’s workflow, interacting with it becomes completely automatic. During the first few days of a deprivation study, users will frequently experience UX “phantom limbs.” They will instinctively reach into their pocket for a smartphone app that has been deleted. They will automatically hit the keyboard shortcut for the software they are banned from using, only to stare blankly at an error message.

In their daily logs, participants will record these reflexive moments. These phantom impulses are a goldmine for interaction designers. They map precisely to the deepest grooves of the user’s mental model. If users repeatedly report reaching for a specific “quick-add” button or swiping in a specific direction when the app is gone, you have identified the core behavioral anchors of your product. These are the micro-interactions you must prioritize, protect from unnecessary redesigns, and make as frictionless as computationally possible.

Reflexive attempts to use something that has been removed are great data. (GPT-Images-2)

These reflexes are also a warning against redesigns that measure success only by first-use comprehension, as frequently studied in standard usability testing. A new interface can test beautifully in a lab and still destroy value if it breaks a high-frequency reflex that experienced users have spent years automating. For mature products, learnability is not enough. The interface must preserve procedural memory.

However, reflexive habits also reveal a psychological vulnerability: cognitive deskilling. By removing the tool, you often discover the user has not merely outsourced a task to the software, but has completely lost the underlying skill required to perform it independently. When a user can no longer navigate their own city without a GPS, or structure a coherent paragraph without an AI assistant, the deprivation study measures the exact degree to which your product has atrophied human capability. This forces teams to grapple with an ethical question: is your software elevating human potential, or systematically degrading it to ensure dependency?

2. The Duct-Tape Workarounds (Compensatory Behaviors)

This is perhaps the most valuable data point a deprivation study can generate. When a user’s primary tool is removed, their underlying goal does not magically disappear. To accomplish their tasks, they will string together alternative solutions. I refer to these as “duct-tape workarounds.”

If you take away the irrigation, the vegies establish a workaround system, so clearly they need the water. Irrigation was the tool, water was the goal. (GPT-Images-2)

Imagine you deprive a sales team of their complex, enterprise Customer Relationship Management (CRM) system. How do they track their leads? They do not simply stop selling. Instead, you might find they revert to a chaotic mix of text messages for urgent updates, sticky notes on their monitors for reminders, and personal Excel spreadsheets for forecasting.

By analyzing these messy workarounds, you see your product’s value disaggregated into its raw components. You learn that your CRM wasn’t just “database software.” It was an anxiety-management tool, a daily task prioritizer, and an asynchronous communication channel.

Furthermore, if the user’s duct-tape workaround is actually easier, faster, or less frustrating than using your official product, you have discovered a critical design failure. If taking away your bloated project management tool results in users happily and efficiently managing their tasks with a simple physical notepad, your software is vastly over-engineered. The interaction cost of your UI has exceeded the value of the tool.

Workarounds should also be analyzed for who pays the cost. Sometimes the deprived participant appears to cope, but only because an assistant, teammate, customer, or family member absorbs the missing coordination work. The product’s true value may reside in preventing these invisible cost transfers. A deprivation study should therefore record not only the participant’s workaround, but also the bystanders who were recruited into it.

You may discover that several bystanders, besides the participant in a deprivation study, are key elements in the workflow and are impacted by the deprivation. (GPT-Images-2)

Such broader observation exposes the “blast radius” of the missing product. Depriving a single user of an enterprise collaboration tool doesn’t just inconvenience them; it creates collateral friction for their colleagues, managers, or downstream systems who silently rely on their data exhaust. This network contagion provides an incredibly rare metric: the measurement of a product’s invisible ecosystem value, revealing that the true anchor of your software might sometimes be the secondary users who never even log in.

3. The Discovery of “Ghost Features” (The Apathy of the Unused Function)

This is often the most sobering and humbling finding for product teams. You will have dedicated hundreds of engineering hours and massive budgets to developing a complex feature, only to find that during the deprivation period, not a single user laments its absence.

Self-reported survey data might tell you a feature is a “nice to have.” Deprivation data tells you whether it is a ghost feature. If users had a realistic opportunity to need a feature during the study and did not miss it, treat that feature as a candidate for demotion, consolidation, or deletion.

However, absence of complaint must be interpreted against opportunity to need. Some features are insurance features: rarely used, but crucial when a crisis occurs. To classify a feature as a true ghost, compare deprivation silence against baseline frequency, task exposure during the study window, and the severity of the consequences if the feature were needed. Ordinary unused features should be pruned aggressively. Rare but high-consequence features should be demoted, simplified, or moved out of the main path, not blindly deleted.

Featuritis is a common affliction of modern software which often comes with a plentitude of features users will never miss. (GPT-Images-2)

4. The Relief Phenomenon (Identifying Hostile UX)

Not all deprivation is painful. One of the most shocking findings you can uncover in a deprivation study is user relief.

When researchers deprive users of certain social media platforms, notification-heavy news aggregators, or overly aggressive enterprise monitoring dashboards, participants often report a significant decrease in anxiety, an increase in deep-focus productivity, and a general sense of calm. They do not miss the product; they are profoundly glad it is gone.

Users may discover that they are happy enough after locking away your technology. (GPT-Images-2)

If a deprivation study reveals that your product’s absence causes relief rather than frustration, you have a massive strategic business problem. Your product is likely surviving purely on dark patterns, addictive behavioral loops, or rigid corporate mandates rather than genuine, user-centric value. Long-term, products that cause relief upon removal are highly vulnerable to catastrophic churn. For interaction designers, this is a glaring red flag that the system’s notification strategy, information density, or cognitive load is completely overwhelming the user and creating a hostile user experience.

Relief is not the opposite of value. It is often a sign that value and harm have been bundled together. Users may still need the product’s record keeping, search, coordination, or accountability layer while hating its interruption layer. The design response should not be to kill the product. It should be to separate the utility users need from the coercive surface that exhausts them.

In enterprise settings, relief can also point to a much deeper organizational pathology: the tool was acting as a proxy for broken processes. If depriving a team of a convoluted project management app suddenly forces them to have a highly effective, five-minute morning stand-up, the software wasn’t solving a communication problem; it was allowing management to avoid fixing one. Deprivation studies don’t just debug software; they debug company culture.

5. Identity and Meaning

Some products become part of how users see themselves. A professional tool may make a user feel competent. A fitness device may make discipline visible. A creative platform may support self-expression. A social app may maintain belonging. When deprived, users may experience not just inconvenience but a change in self-perception. These findings should be handled carefully because they are often less about interface mechanics and more about product positioning, onboarding, community design, and long-term engagement.

Guidelines for Conducting the Study

Running a deprivation study requires meticulous planning. You are asking users to do something highly unnatural and disruptive to their daily lives. If you execute the methodology poorly, users will simply cheat by breaking the deprivation rules, or they will drop out entirely, rendering your data corrupted and useless.

What if user research was a role-playing adventure game (RPG)? This is your player class card with all the equipment and skills to conduct a deprivation study. (GPT-Images-2)

Here is pragmatic, step-by-step advice for setting up and executing a valid deprivation study.

Step 1: Recruiting and Incentivization (The “Pay to Quit” Model)

Recruiting for a deprivation study is notoriously difficult. You are not asking for sixty minutes of their time in a comfortable usability lab; you are actively sabotaging their daily routine.

First, strictly screen for active, habitual users by verifying high-frequency usage through backend analytics or screener surveys; depriving occasional users yields no meaningful data.

Second, you must pay them significantly more than standard usability participants. A typical one-hour lab study might require a $100 incentive. A two-week deprivation study might require $500 to $1,000 per user, depending on the severity of the deprivation and the professional level of the participant. You are paying them a premium for their intense inconvenience and to guarantee their strict compliance.

The harsher the study, the higher the participant incentive. (GPT-Images-2)

I highly recommend a staggered payment structure: a small payment for the baseline phase, a moderate payment during the deprivation, and a massive completion bonus at the very end to prevent dropouts on Day 4 when the frustration peaks. You only need 12 to 15 highly compliant users to reach qualitative saturation.

Step 2: Establishing the Baseline (Pre-Deprivation)

You cannot accurately measure the impact of absence without first measuring the reality of presence. Before the deprivation begins, run a baseline phase for 3 to 7 days.

During this phase, instruct the user to behave normally. However, have them meticulously log their usage of the target product. Use a lightweight diary study format. Every time they use the product, they should note their goal, the time taken, and their emotional state. Better yet, with explicit user consent, use screen-tracking or passive analytics software to objectively log their baseline usage. This provides a factual, empirical record of their habits to compare against their deprived state.

Before you can measure absence, you must first document presence, but without turning the baseline into the intervention. (GPT-Images-2)

Step 3: The Deprivation Period and Data Collection

The length of the deprivation period is a delicate balance. If it is too short (e.g., 24 hours), the user treats it as a novelty or simply delays their tasks until the study is over. You will only capture initial shock, not adaptation. If it is too long (e.g., a full month), attrition rates will skyrocket, and users will permanently transition to a competitor.

• The sweet spot for most digital products is 7 to 14 days.

The first three days will be characterized by intense frustration, those “phantom limb” impulses, and initial resistance. By days four through seven, users will have actively established their duct-tape workarounds. By week two, you will see which workarounds are sustainable and which ones are causing compounding, cumulative pain.

Track three time-based measures. First, replacement half-life: how long it takes users to settle on a stable substitute. Second, backlog elasticity: how much work they postpone because only your product handles it well. Third, relapse pressure: the point at which participants try to break the deprivation rules. These measures make a qualitative method easier to explain to executives because they translate lived friction into observable behavioral thresholds.

Deprivation studies produce rich data. (GPT-Images-2)

During this period, data collection must be continuous but low-friction. Do not rely solely on a single interview at the end of the two weeks. Human memory will blur the micro-frustrations experienced on Day 2. You must capture the data in situ. Use the Experience Sampling Method (ESM). Ping the users via SMS or a dedicated diary app two to three times a day at random intervals. Ask them simple questions:

1. Did you encounter a situation today where you desperately wanted to use the deprived product?

2. What specific task were you trying to achieve?

3. What alternative method did you use instead?

4. Rate the extra effort required by your alternative solution on a scale of 1 to 7.

Ask users to keep a log or diary of deprivation events and thoughts, but don’t depend on them to systematically do so without reminders. (GPT-Images-2)

Step 4: The Post-Deprivation Reunion and Debrief

The exact moment the user is reunited with the product is incredibly illuminating. Do not let this moment pass unobserved. Have the user log back into the product while you watch them, either in person or via remote screen-share.

The first actions they take upon returning to the product are highly revealing. Treat the reunion as a priority sort, not a spectacle. The first item opened, the first setting restored, the first search executed, and the first message sent are all behavioral votes. They show which part of the product users trusted enough to save for the official tool, and which parts they were willing to let substitutes handle. The actions taken in the first ten minutes of reunion point directly to your product’s most vital, irreplaceable core utility.

When people resume the technology, which feature do they turn to first? (GPT-Images-2)

Following the reunion, conduct a comprehensive, 60-minute qualitative exit interview. Review their diary entries with them and ask them to elaborate on the specific workarounds they documented.

Turning Findings into Actionable Design Recommendations

Research without action is merely expensive corporate trivia. It is a waste of resources. The ultimate goal of enduring the logistical headache of a deprivation study is not to write a fascinating sociological white paper; it is to force profitable improvements to your product’s user interface.

Translating the messy, qualitative, emotional data of a deprivation study into wireframes or a product roadmap requires ruthless pragmatism. Here is how you map your findings to actionable design decisions.

1. Paving the Cow Paths (Formalizing Workarounds)

In urban planning, there is a concept called “paving the cow paths.” Instead of laying down concrete sidewalks where architects think people will walk, planners wait to see where the grass is naturally worn down by foot traffic, and then they pave over those dirt paths.

In a deprivation study, the users’ duct-tape workarounds are your cow paths. Analyze them carefully. If multiple users in your study created the exact same manual workaround to survive without your product, you have discovered a critical missing feature in your current design.

What people do when left to their own devices can serve as a good starting point for specifying new products or features. But don’t replicate unnecessary complications that are an artifact of technology deprivation and not a true need. (GPT-Images-2)

But pave only the desire path, not the dirt. A workaround often contains accidental complexity because the user lacks the right tool. The product team should extract the intent from the workaround, then design the simplest native path to that intent. Copying the workaround literally risks institutionalizing the very friction the study exposed.

Actionable Recommendation: Integrate the user’s workaround directly into your product natively. Stop fighting user behavior and start facilitating it. For example, if you deprived users of a complex accounting software, and they all reverted to typing raw numbers into a plain-text document because your official form fields were too restrictive, you must build a “plain-text quick-entry” module natively into your UI. Lower the interaction cost of their natural, observed behavior.

2. Elevating the “Invisible Core” (Restructuring Information Architecture)

Your study will definitively reveal the features users missed the most: the ones that caused the most acute, verifiable pain when removed. Frequently, usability audits discover that the feature users value the absolute most is buried three clicks deep in a settings menu, while the home screen is dominated by marketing fluff or useless secondary widgets.

If users’ diary entries were primarily focused on lamenting the loss of one specific function, that function is your product’s true center of gravity.

Actionable Recommendation: Radically restructure your Information Architecture (IA). Elevate the core function. If a deprivation study proves that users missed your advanced search capability vastly more than your curated content feed, the search bar must become the most prominent, visually weighted element on the screen. The interaction cost to initiate this core utility must be reduced to an absolute minimum, ideally a single click or tap from anywhere in the system. The UI must reflect the user’s hierarchy of needs.

When deprivation research demonstrates special user love for certain features, that’s a hint to redesign to make those features more prominent and faster to reach. (GPT-Images-2)

3. Ruthless Pruning (The Strategy of Subtraction)

As established, your interface is likely cluttered with features that users do not use and do not care about. We know from Hick’s Law that the time it takes for a user to make a decision increases logarithmically with the number and complexity of choices presented to them. Every unused feature on your screen degrades the usability of the essential features by creating visual noise.

Review your deprivation data for the ghost features: the ones no one mentioned missing.

Cutting features is one of the best ways to increase usability. Less is More. (GPT-Images-2)

Actionable Recommendation: Deprecate them. Remove the dead weight to clean up the graphical user interface. If internal political pressure from stakeholders prevents outright deletion, aggressively demote the features. Move them out of the primary navigation and bury them in an “Advanced Settings” menu. Use the empirical deprivation data as an impenetrable shield against stakeholders who claim, “But users will revolt if we remove this widget!” You have documented proof that they will not even notice. Good usability is often defined by what you choose not to show on the screen.

4. Redesigning Onboarding for “Time-to-Value”

Most software onboarding flows are atrocious. They force the new user through a tedious, multi-screen carousel explaining the mechanical functionality of every minor icon in the interface. They fail entirely to teach the user why they should care about the product.

Deprivation studies give you the ultimate cheat sheet for onboarding. The study tells you exactly what long-term, habituated users value. New users do not yet know this value; they must be guided to it immediately.

Actionable Recommendation: Rewrite your onboarding copy and restructure your tutorials based exclusively on the insights from the deprivation study. Strip away everything else. Do not teach new users about secondary features. Accelerate the user’s journey to the core utility; the exact feature that experienced users rushed to use during the Reunion phase. Teach them the one or two core behaviors that will form an unbreakable habit, and let them discover the rest later. Optimize strictly for rapid “Time-to-Value.”

You can drastically cut the overhead of onboarding and other instructions by focusing them on the features users need the most, as demonstrated by deprivation research. (GPT-Images-2)

5. Aligning Marketing Strategy with Pragmatic Reality

Usability does not stop at the edge of the screen; it extends to the words you use to describe the product on your landing pages. Often, a company’s marketing department is selling an idealized fantasy, while the user is buying a purely pragmatic reality.

Actionable Recommendation: Rewrite your marketing copy, empty states, and error messages using the exact vocabulary your participants utilized in their diary logs. If your marketing proudly proclaims that your app “Synergizes Enterprise Communication Paradigms,” but your deprivation study revealed that users solely value the tool because it “stops me from losing track of PDF attachments in my inbox,” you must change your messaging. Speak to the pragmatic, observed reality of the user’s pain points. In user experience, clarity will always outperform clever marketing jargon.

Marketese rarely speak to users’ main needs. A deprivation study identifies the users main pain points that your product solves. Use this information to transform product communication into plain language that’s explicit about the product benefits. This is a major recommendation for digital content anyway. (GPT-Images-2)

Case Studies: Web and Smartphone Deprivation

Web Deprivation: From Utility Loss to Conditioned Withdrawal

Early web-deprivation research from 2004 through 2015 established a chilling fact: the desktop internet had already transitioned from a convenience into non-negotiable cognitive infrastructure.

The 2004 Yahoo!/OMD study locked 28 people offline for two weeks. Withdrawal was immediate and brutal: participants reported isolation, frustration, and a startling loss of basic analog skills. They had forgotten how to use a phone book, consult a newspaper, or even call a customer-service line. The browser had become the master interface for daily logistics, and removing it felt less like a detox and more like an amputation of everyday competence.

Web deprivation studies required users to relearn behaviors they had forgotten in favor of more convenient online access to the world. (GPT-Images-2)

By 2010, “The World Unplugged” study quantified the emotional toll with 1,000 students across five continents attempting a mere 24 hours without any media. Most failed. Those who endured generated 3.5 times more negative descriptors than positive ones, freely using the language of clinical pathology: anxious, empty, desperate, dead. The core pain came not from missing facts but from losing what researchers called “affective awareness”: the continuous background hum of knowing your social circle’s status without interacting. The web had grown invisible until severed, and then the void induced panic.

Subsequent internet-deprivation research added two more useful signals: conditioning and physiological arousal. Osborne and colleagues’ 100-participant experiment found that higher-problem internet users who had been deprived for four hours showed lower mood, greater anxiety after web cessation, and a conditioned shift toward colors prominent on the websites they had just visited. A separate Swansea/Milan study with 144 participants found increased heart rate and blood pressure after internet use ended among users with high problematic-use scores. These findings are most useful for UX research not because every heavy user is an addict, but because they show how technology can become a negative reinforcer: the user returns not for pleasure, but to stop discomfort.

Smartphone Deprivation: Neurological Rewiring and Attentional Recovery

With the smartphone, the dependency model escalated from episodic task utility to continuous ambient attachment, and the deprivation research of the 2020s exposes the startling cognitive and neurological toll.

In 2025, Ward and colleagues performed the definitive experiment: 467 adults disabled mobile internet on their phones for two weeks while keeping calls, texts, and desktop web intact. The results were so dramatic they merit reading twice. 91% of participants improved on at least one mental-health or cognitive metric. Depression symptoms fell by an effect size larger than that observed in multiple antidepressant trials. Sustained attention, the capacity every productivity method tries to restore, improved by an amount equivalent to reversing a full decade of age-related cognitive decline. People didn’t substitute other screens; they reclaimed offline life, exercising, socializing face-to-face, and sleeping more, with benefits growing stronger day by day. The constant mobile data drip was not neutral; it was actively corrosive to focus and well-being.

Deprived of smartphone use, participants reclaimed offline life and their mental health improved. (GPT-Images-2)

The compliance data from this study may be as important as the outcome data. Only part of the sample fully carried out the restriction, which means the study did not merely measure what happens when mobile internet is blocked. It also measured how hard it is to keep mobile internet blocked. In deprivation research, noncompliance is not just noise. It is behavioral evidence that the intervention is fighting a real dependency.

Neuroimaging work from the same year by Schmitgen’s team dug into the brain mechanisms. After just 72 hours without a smartphone, young adults showed altered activation in the anterior cingulate cortex, insula, and striatum when shown pictures of phones: regions at the heart of dopamine-driven craving and habit circuits. The neural signatures matched those of substance-use disorder cue reactivity. The phone had become a neurobiological trigger.

2024 trials that capped daily screen time at two hours produced consistent small-to-medium effect-size gains in sleep quality, stress reduction, and mood, proving the benefits are neither transient nor placebo. Smartphone deprivation does not simply cause irritation; it triggers a withdrawal syndrome in the brain’s reward architecture, and sustained removal allows measurable recovery. Effect sizes ranged from small to medium but were durable rather than transient.

Across all three studies, the signal is consistent: smartphones function as neurobiological triggers whose continuous dopaminergic stimulation degrades cognition, mood, and sleep. Remove the trigger for sustained periods, and the brain rebuilds baseline functional capacity remarkably quickly.

Smartphones consume attention even when not used. (GPT-Images-2)

The Dependency Shift: From Tools to Ambient Parasites

The synthesis of two decades of research reveals a clear, qualitative shift. Desktop-web dependency was task-centric: users suffered when they could not accomplish specific information goals, as though a library and post office had been simultaneously demolished. Frustration was logistical.

Smartphone dependency is an ambient attack on the brain’s attentional control system. The device’s constant bodily presence has conditioned the brain to anticipate variable rewards at any moment, where even a powered-off phone consumes cognitive resources by forcing the user to suppress the urge to check.

Classic deprivation studies of desktop web access and smartphones revealed high dependency on these two technologies, but in different ways. (GPT-Images-2)

The withdrawal profile underscores the change. Early web studies showed yearning for functional access; smartphone studies yield panic over lost affective awareness and trigger addiction-style neural craving.

The recovery data offers optimism through a measurable rebound though not a full cure. When mobile internet is reduced for sustained periods, many users regain time, attention, sleep, social connection, and perceived control. The product lesson is not that every smartphone interaction is pathological; it is that continuous access can turn a useful tool into a standing claim on the user’s mind.

Conclusion: The ROI of Subtraction

When you propose a deprivation study, you will face intense pushback from product managers and executives. The idea of intentionally stopping paying users from engaging with your product sounds counterintuitive, even dangerous, to short-term business metrics. They will argue it will hurt Daily Active Users (DAU) or cause churn.

You must counter this skepticism with the logic of Return on Investment (ROI).

The technology industry wastes billions of dollars every year designing, coding, testing, and maintaining features that serve absolutely no human need. Development teams spend months agonizing over the micro-interactions of a tool that users would gladly trade for a simple spreadsheet. We suffer from a severe addition bias: we assume the solution to any usability problem is to add more UI.

The most expensive usability mistake is adding even more features when subtraction would create more user value. (GPT-Images-2)

The most expensive mistake a software company can make is to write code for a feature that nobody wants. The second most expensive mistake is to bury a critically important feature beneath layers of convoluted navigation.

Standard usability testing will tell you if a user can navigate your interface. A deprivation study tells you if they should. It forces a stark confrontation with the truth of your product’s utility. Deprivation studies enforce a strict discipline of subtraction. They act as a brutal reality check, stripping away the superficial layers of user preference, stakeholder politics, and industry fads to expose the bedrock of actual user need.

Yes, the methodology is demanding. It requires meaningful incentives, disciplined fieldwork, and rigorous qualitative analysis. But the ROI can be unusually high because the method attacks the two most expensive product mistakes: building features nobody needs and hiding the few features users cannot live without.

If you want to know how users navigate your interface, sit them in a lab and watch them click. But if you want to know if your interface deserves to exist in the real world, take it away from users. The resulting silence, or the ensuing chaos, will tell you everything you need to know.

Would anybody watch an anime movie about a deprivation study? There could be plenty of drama. My Japanese friends: please make one! (GPT-Images-2)

Watch my short explainer video about deprivation studies (YouTube, 6 min.)

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,201 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: Claude Design should be used differently than canvas-based design tools | AI conducts user testing and can now see what’s happening on the screen | For the first time, more than 50% of employed Americans use AI at work | OpenAI recaptures leadership in AI image generation with new, highly accurate model | Two upgraded AI models, GPT 5.5 and DeepSeek 4

UX Roundup for April 27, 2026 (GPT-Images-2)

Claude Design ≠ Smart Figma

Claude Design is Anthropic’s new AI-native design environment to generate and iteratively refine UI designs, slide decks, and interactive prototypes by conversing with Claude rather than drawing everything manually. It runs on Anthropic’s latest vision model (Claude Opus 4.7).

Claude Design is a browser-based canvas paired with a chat pane where you describe what you want (e.g., “mobile onboarding flow for a budgeting app for retirees”) and get a structured layout, styling, and content automatically. You then refine that output using natural language, inline comments, or direct edits on the canvas, with Claude updating the full design systemically.

Early feedback emphasizes the benefits of the designs aligning with your established design system, and it’s strongly recommended to upload your complete design system to Claude Design before using it to produce any designs. Claude Design can crawl your codebase and design files to automatically infer a design system if you don’t have an explicit design system already defined.

Ryan Mather, who is a designer at Anthropic, posted his initial advice for using Claude Design. Three points that stood out to me:

• Do spend time upfront to set up your design system in Claude Design and upload the core screens from your existing design.

• Iterate with your engineers live: you can now get live protypes in real time, making for deeper discussions about proposed new features.

• Don’t use chat, but point to the places in the design you want to change. This is much easier than a verbal description of changes.

The overarching insight from Mather is that Claude Design is a different beast than traditional design tools that are based on laying out individual UI screens one at a time. The new tool redefines the job of designing.

The primary artifact shifts from the screen to the system. Traditional canvas tools treat the screen as the unit of work. You open a frame, you drag components onto it, you finish a screen, you start another. Claude Design inverts this: the design system is the primary artifact, and individual screens are ephemeral outputs of that system applied to a brief. This is why Mather’s advice about uploading your design system first isn’t housekeeping; it is the work. If your design system is thin or inconsistent, your outputs will be thin and inconsistent, no matter how cleverly you prompt. If your system is rich and opinionated, Claude has more constraints to push against, and the generated designs feel like they actually belong to your product rather than to the model’s statistical average of “good SaaS UI.”

The cost curve of exploration inverts. In Figma, producing three credible variants of an onboarding flow takes hours, so designers tend to commit to a direction early and refine it. In Claude Design, generating five directions costs about as much as describing them, so the bottleneck shifts from production to evaluation. This sounds like pure upside, but many designers report feeling overwhelmed at first: they haven’t needed to develop a vocabulary for rapidly discriminating between options, because they’ve never had this many options. Taste becomes a throughput constraint rather than an output quality.

Fidelity stops being a ladder. The traditional progression from sketch, wireframe, mockup, prototype, to released product exists because each rung costs more than the one below it, so you do cheap work first to de-risk expensive work later. Claude Design collapses this ladder by making high-fidelity interactive prototypes nearly free. This is liberating, but it’s also the biggest trap in the tool. You skip the “low-fi thinking” stage where you’re forced to wrestle with information architecture and flow before getting seduced by visual polish. The discipline of deliberately working at low fidelity, even when you don’t have to, is an emerging best practice.

Why pointing beats describing. Mather’s advice to point at the canvas rather than chat deserves unpacking. Natural language is excellent at conveying intent (“make this feel more trustworthy,” “reduce the density”) but terrible at conveying location (“the third card in the second row, but only the variant that shows when the user is logged out”). Pointing resolves the referential ambiguity that chat handles poorly. The cleanest workflow treats the canvas as the “where” and the chat as the “what to change.” Mixing them (e.g., typing “make the blue button on the second screen slightly smaller”) is where Claude Design sessions tend to go off the rails.

The critique loop changes shape. In Figma, feedback flows from reviewer to designer, and the designer adds value by interpreting vague feedback into concrete changes. In Claude Design, reviewers can talk to the system directly. This weakens the designer’s role as a feedback-translator and strengthens his or her role as constraint-setter: the person who defined the design system, wrote the brief, and chose the initial direction. Designers who have built their careers on being “the one who can turn ‘make it pop’ into actual pixels” will find that leverage shifting.

Engineers become collaborators, not stakeholders. Mather’s point about iterating with engineers live is another 180° flip in the UX workflow. In traditional workflows, engineers arrive after designs are “done” and push back on feasibility, creating expensive rework. When engineers can propose changes directly in the design environment to test “what if this list virtualized differently” or “what happens to the layout at this breakpoint,” they’re co-designing rather than gatekeeping. The design/engineering boundary gets productively fuzzier.

The failure modes are different, and worth learning to spot. Figma designs fail by being inconsistent, off-brand, or physically impossible to implement. Claude Design outputs fail differently: they tend to be internally consistent but generic, exhibiting a kind of aesthetic mode collapse toward the average of well-designed products. They can be confidently wrong about things that don’t translate to natural language, such as microinteractions, timing, spatial rhythm, the specific way a particular element should feel under the finger. (Though all of these will obviously get better in the next release.) And they can feel over-optimized, as though every decision has been justified, which paradoxically makes them feel soulless. Learning to recognize and correct for these failure modes is the new craft.

The competency stack shifts. Traditional design workflows reward visual craft as a production skill. Claude Design rewards it as a judgment skill: you still need to know what “good” looks like, but you’re using that knowledge to evaluate output rather than to produce it. Alongside this, three competencies become disproportionately valuable: clear articulation of intent and constraints, strong design system literacy (knowing what’s worth codifying and what should stay flexible), and curatorial stamina (the ability to evaluate a lot of options without defaulting to the first acceptable one). None of this makes designers obsolete. It makes the median designer’s job look more like what senior designers and design directors already do: setting direction, defining standards, and exercising judgment over output they didn’t personally produce.

All of these shifts mean that Claude Design moves designers up the abstraction ladder. You’re no longer the person placing rectangles; you’re the person defining what rectangles should exist and why. That’s a bigger change than any feature list conveys, and it’s why teams who treat Claude Design as “Figma, but I type instead of drag” tend to underwhelm themselves with it.

For the sake of variation, I replaced my usual Silicon Valley product team with one based in Tokyo to bring you this comic strip about the new design process:

(Nano Banana 2)

Even though I refer to Claude Design in the above analysis of how systems-capable AI design tool change the design process, I expect that the implications of the forthcoming competing tools from the likes of Google and OpenAI will be much the same. AI is super-competitive, and better tools will likely emerge soon.

10 Worst UI Annoyances

I made a new music video about the top-10 most annoying UI designs (YouTube, 3 min.) For more detail about each of these annoying design mistakes, read my in-depth article.

10 UI design mistakes that annoy users (and thus cost sales): my newest song. (Nano Banana 2)

I have become dissatisfied with avatar animation recently, because it hasn’t kept up with advances in AI video. I therefore decided to skip the singing avatar for this video and purely rely on B-roll clips that illustrate the lyrics, even though this is more work, because of the need to generate so many separate clips. This video needed 32 generations for 3 minutes and 32 seconds of music. (In total, the video has 66 cuts, because many of the generations used the multi-shot capability of Seedance 2.)

I used Seedance 2 for most of the clips, because it is currently the best video model. (I made a few clips with Kling 3 and one with Veo 3.1, which is starting to show its age but still has strengths.) Most of my generations were pure text-to-video, but I also made frequent use of omni-reference generations based on still images made with Nano Banana 2.

I am the most pleased with the clip of the design manager lovingly polishing her design award for cool design and the gorilla suffering fat-finger problem while using his smartphone. (The gorilla clip benefited from upscaling to 4K with Topaz, but mostly I think the upscaling was a waste of credits.)

On the other hand, I was disappointed with the clip of the password judge (for which I had higher hopes) and the clip of menu commands hiding behind a burger menu.

Which clips did you like the most or the least? Let me know in the comments.

Compared to another avatar-less music video I made recently, Hamlet: The Music Video (YouTube, 4 min.), I feel that Hamlet was more successful. UI Annoyances compiled a series of unrelated clips made in a wide variety of animation styles and with no pervasive characters, which gives it a less cohesive feel than Hamlet, which is a storytelling video using a small set of characters throughout and with all the clips made in the same animation style. (Even though Seedance didn’t perfectly obey my style references for all the clips: note how Hamlet’s animation style drifts when he’s pretending to be mad — possibly because the character reference was different for those two scenes.)

Combining video clips without a unifying avatar character may be best suited for videos with a single storyline, like a Shakespeare play, as opposed to montage-style content like a “top-10” list. (Nano Banana 2)

AI User Research Service Adds Vision

The AI-powered user research service Outset has added the ability for the AI to watch what users are doing on-screen or what they are doing in the physical environment, such as in-store shopping, for studies.

AI is gaining capabilities to perform more steps in usability testing. (Nano Banana Pro)

For a few years, AI user research has been able to listen to test participants’ comments and analyze what they say. But anyone who has ever done user testing knows there is often a distinct difference between what users say and what they do. An AI analysis based solely on users’ verbal statements will always be second-best.

I am very happy to see the “do” side of user research being tackled by AI research services. Admittedly, the original LLM-based AI was mostly suited to language/utterance analysis, but now AI is becoming multimodal and adding world models, and user research should benefit! For example, Meta’s new Muse Spark AI model emphasizes multi-modal and visual reasoning, and Google’s Gemini Pro has also been getting stronger in moving beyond pure language processing.

AI is improving its vision capabilities, which is hugely useful for understanding and correctly analyzing what’s happening in a usability test. (Nano Banana 2)

This new ability for AI to both watch users and listen to their think-aloud commentary will make unmoderated user testing much cheaper and thus more useful. Two ways AI analysis helps: First, if a human UX expert doesn’t have to watch the recordings, it becomes feasible to test with more users, for example, to get true international coverage. Second, when testing is cheap, more will be done. Human-powered user research will always be rare, no matter how much I have been pushing fast “discount” usability studies. But when all you have to do is push a button to gain insights into how your customers use your design, then it becomes feasible to test many more design ideas. A broader range of ideas again means a higher-quality one you pick for implementation.

Old-school usability folks like myself are nostalgic for the days when we sat with the users in the lab and moderated test sessions in person. But this was much more expensive than having an AI conduct the test and analyze the results, so AI-driven user research is clearly the future. (Nano Banana 2)

Despite this optimistic view, it’s probably prudent for a human UX expert to review a few of the session recordings rather than rely only on an AI analysis of what happened during the test. It might be unfair of me to say this, because I don’t know how good Outset’s AI is at usability analysis. Maybe it’s perfect. But more likely, it still has weaknesses and will improve over the next year or two, as they gather more training data. (I believe firmly in the usability scaling law: as AI collects more user data, it becomes better at judging design quality and analyzing usability problems.)

In the beginning, I recommend watching a few session recordings manually, to check up on the AI’s usability analysis. (Nano Banana 2)

Bottom line: more AI = cheaper studies = more user research = better usability. (Nano Banana 2)

More Than 50% Of Employed Americans Use AI At Work

For the first time, more people use AI than not at work in the United States, confirming how AI adaptation is growing faster than any previous technology. This according to a Gallup survey of 23,717 U.S. employees conducted in February 2026. (Since this is two-months-old data, the current share of U.S. workers using AI is probably 52%, and not the reported 50%. Usage grew by 3 percentage points since the previous quarter, corresponding to roughly one percentage point per month.)

We passed the tipping point where more people use AI than not in American businesses. (NotebookLM)

Even though 50% of employed respondents reported using AI at work in February, only 13% said that they did so daily. This means that about three-quarters of AI users are still stuck at intermittent use, as opposed to the intensive, workflow-changing AI use that’s required to transform the economy and explode our living standards.

Another data point from the Gallup report confirms this conclusion: While 65% of employees in AIU-adopting organizations say that AI has improved their personal productivity, only 31% (less than half) say that AI has transformed how work gets done in their company. I even suspect that this is an exaggeration since most people don’t understand the potential for AI to fundamentally transform workflows. (That’s why we need people like you, Dear Reader, to help companies redesign for AI.)

While people report higher individual productivity, this is insufficient to truly lift the economy, if companies simply use AI to (metaphorically) operate the fax machine. Full workflow transformation is needed. (NotebookLM)

As you know, there’s much gnashing of teeth in the press about possible job losses due to AI. However, Gallup’s data doesn’t support this negative attitude. It is true that AI-adopting organizations have more workforce change than organizations not adopting AI: 57% of AI-using companies are expanding or shrinking their workforce, compared with only 44% of non-adopting companies. (This simply confirms that any company that’s still not on the AI bandwagon must be incredibly sluggish and reluctant to change.)

More interesting, of those AI-adopting companies that have changed the size of their workforce, 60% increased their staffing, whereas only 40% decreased staffing. Of course, many other things influence staff numbers that merely AI use, especially with the relatively modest AI use most companies still experience. However, this number certainly refuses the fear that AI is a ruthless job killer.

While AI-using companies are more dynamic than non-users and therefore have more staffing-level adjustments, they bring in more new hires than they fire old staff. (NotebookLM)

A final interesting observation is that people are more positive toward AI the higher their level. Respondents answering “extremely positive” or “somewhat positive,” by company rank:

• Leaders: 71%

• Managers: 59%

• Individual contributors: 54%

New data from Gallup documents the rapidly increasing use of AI in American business. (NotebookLM)

A survey by a competing company confirms Gallup’s findings, which vastly increases the credibility of the findings. An Ipsos poll released on April 13, 2026 (and conducted March 3–6, 2026) surveyed 2,021 U.S. adults about their use of AI tools. Half of respondents reported using at least one AI service, with ChatGPT being the most widely used; common tasks include information lookup, writing or editing text, and brainstorming. Among working adults who use AI, 51 % leverage it on the job.

50% of Americans say they have used an AI service in the past week. ChatGPT leads utilization at 31%, followed by Google Gemini (21%), Microsoft Copilot (11%), Meta AI (8%), Grok (5%), and Claude (3%). The main surprise to me is that Meta scores so high, which demonstrates the advantage they get from distribution to existing users of Facebook/Instagram/WhatsApp.

A second surprise is that Claude scores so low, since it’s often considered to be one of the “Big-3” frontier models. It’s good to be reminded that the broad public are not AI influencers or fanatics in any way and don’t follow the usage patterns of the leading (and bleeding) edge. Confirming this point, among respondents who said that they used AI during the last week, most (62%) said that they only used it “a little” (for one or two quick tasks). Only 6% used AI “a lot” (used repeatedly or relied on it heavily throughout the day).

AI has “crossed the chasm” and gone mainstream to the extent that more people use it than not. But most people’s usage is still light, confirming that majority users are very different from that of innovators and early adopters like you and me. (GPT-Images-2)

What people do with AI: very pragmatic uses dominate. Note how many people have used image generation: everybody can use a pretty or ironic picture. Agents are still emerging. (GPT-Images-2)

GPT-Images-2: Thinking Pictures

OpenAI launched its new GPT-Images-2 model on April 21 and probably reclaimed the throne as best image model, at least for complex images with extensive text. The original GPT-Image-1 model from March 2025 was a revolution, being the first image model that was integrated with a general AI language model. (In contrast, GPT-Image-1.5 from December 2025 was a grave disappointment not reaching anywhere near Nano Banana Pro’s capabilities.)

As a first experiment, I asked GPT-Images-2 to draw page 2 from the comic strip I had previously made with Nano Banana 2 about the changing design process with Claude Design. (See that full comic strip earlier in this newsletter.) I gave the image model the full text prompt (which specifies the action, setting, captions, and speech bubbles) as well as the character reference sheets plus the fully rendered pages 3 and 4 as style references.

(GPT-Images-2)

In this simple example, GPT-Images-2 defeated Nano Banana 2 with prettier artwork and drawings that more closely report the underlying story. For example, the five screen designs that Sakura (the designer) generates in the middle left panel are more interesting examples of design variations. And in the lower left panel, Hina’s (the researcher) low-fidelity prototype appears on the tablet she’s holding instead of being on a separate rendering. (On the other hand, Nano Banana 2 showed Hina holding a normal-sized tablet, whereas GPT-Images-2 made the tablet unrealistically big to make room for a legible prototype.)

Simply drawing a pretty comic strip from a given storybook is a poor test of a thinking image model. What’s great about GPT-Images-2 is its ability to reason about visuals before drawing them. As a second example, I asked both GPT-Images-2 and Nano Banana 2 to draw an infographic about my website, www.uxtigers.com. First, the older model, Nano Banana 2:

(Nano Banana 2)

(GPT-Images-2)

Again, I would say that GPT-Images-2 drew prettier visuals, whereas Nano Banana 2 used an overly generic AI-infographic style. GPT also rendered my logo correctly, whereas Nano Banana 2 only showed small cute tigers but didn’t include the actual website logo in its infographic. (Instead, it drew me, which is a nice touch for a personal website.)

Nano Banana 2 included substantially more information about the website, including specific examples of articles within the main themes it identified. GPT-Images-2 only included generic information, such as the names of the themes (though it accurately identified more themes). It also wastefully included the statement “a resource for the global UX community” twice, with different visuals.

I asked GPT-Images-2 to include more information about recent themes I have covered, and it did so very nicely:

(GPT-Images-2. I wonder whether the lone wanderer on the mountain peak at the very bottom is a reference to the image I ran in UX Roundup for March 16, 2026, which again was my homage to German Romantic painter Caspar David Friedrich’s famous painting “Wanderer above the Sea of Fog” [original title: Der Wanderer über dem Nebelmeer], painted around 1817.)

I still think this is a pretty infographic, but an overly dense layout. I did ask for more information, which will inherently make any visualization denser, but I think GPT could have removed or condensed some elements. (For example, it still mentions the global community twice and has superfluous paw marks on two of the three bottom items.)

Infographic prompt credit Marcio Lima 利真 マルシオ.

The mammoth in the infographic above looks much too big relative to the human hunter. In reality, larger mammoths were about twice as tall as an ice age human male. There’s also too much text in this infographic to be easily read at anything less than poster size, and the text is only legible at all because I rendered this image with the 4K version of GPT-Images-2 on Higgsfield.

Revised infographic after asking for less text and the correct size ratio between the human and the mammoth. Even here, the human looks too small, relative to the animal, but at least it’s better. (GPT-Images-2)

Here’s a pair of infographics, both about the same classic paper by Ben Shneiderman about information visualization, but where I instructed the image model to vary the amount of information presented:

Fashion photo shoot of models wearing UX Tigers swag:

(Inspired by “Intent by Discovery: Designing the AI User Experience.”)

(Inspired by “Hamlet Remixed: Using AI to Convert Content into Alternative Formats.”)

As an example of a complex image design, I made a “hidden character” game (prompt credit Kris Kashtanova):

For this kind of complex image, the low resolution of the current model does hurt. (GPT-Images-2)

Rune stone about one of my favorite usability slogans, Less Is More. Somewhat hard to read the text, and that’s after I bumped up the contrast in post-processing. (GPT-Images-2)

To test the new image model’s abilities to draw upon real-time information and to plan consistent visuals across multiple pages, I asked it to draw a 3-page comic strip about how AI influencers and creators have received the new image model:

Great style and character consistency. (GPT-Images-2)

I used GPT-Images-2 to generate all 23 illustrations (plus an OpenGraph image and a social media posting image) for my article last week about 10 steps to build an AI-positive company culture. For each of the 10 steps, I ran pictures showing how they would have played out in the Roman Empire and in a present-day Norwegian company. GPT-Images-2 produced two very consistent image series, as also shown in these concluding images:

Lessons from field research on how to best promote more AI use within a company, as applied to the Roman Empire and in Oslo, Norway. (GPT-Images-2)

Conclusion: GPT-Images-2 is a major step up from the disappointing version 1.5. It is superb at text rendering, with virtually no typos. (In contrast, Nano Banana 2 used to be considered great at text, but frequently misspells words or have the speech bubble tail point to the wrong character in a comic strip. I’m sure Banana 3 will also fix this.) I’ll need more experimentation that goes beyond infographics and comic strips to decide whether I truly like GPT’s aesthetics better than Banana’s, but for now, I’m impressed with my first few tests. GPT-Images-2 is great at producing images based on world knowledge or web knowledge, including real-time information, but not necessarily better than Nano Banana 2.

Some downsides:

• The model is slow. Not quite an overnight-rendering example of truly Slow AI, but a single image often takes about several minutes to render, which is a significant impediment to creation by discovery, since it crimps the number of iterations I’m willing to try to navigate the latent design space for any given idea.

• For now, the model only renders low-resolution images of typically 1086x1448 px for the 3:4 aspect ratio I prefer for comic strips. A 4K option is available via API on external sites like Freepik and Higgsfield (at high rendering cost of 4x that of Nano Banana 2 on Higgsfield and 5x on Freepik), so hopefully that will be coming soon to the ChatGPT mothership. The current resolution is acceptable for web graphics but will limit many other use cases, including the creation of reference images for omni-reference video models, where the final video looks better the more the video generation can extract fine details from the reference images. If you use Seedance 2 or Kling 3, I strongly recommend making 4K reference images.

• Images are often too dark, with low contrast and sometimes beset with speckles, which make photos or detailed drawings less realistic.

PS: Solution to the “Where is Jakob?” image. Eating a hotdog by the hotdog stand in the lower right.

GPT 5.5 Launches

Never a rest if you’re an AI enthusiast. As if the new image model wasn’t enough news from OpenAI, it also launched an upgraded dot-release of its flagship language model, GPT 5.5, on April 23, 2026. This is only 7 weeks after the launch of GPT 5.4 on March 5, 2026. We’re definitely seeing an acceleration in the AI race.

With this little time between releases, it’s to be expected that the differences are less revolutionary than back when we had to go almost a full year before AI was upgraded. The reported benchmark scores are higher, but as I’ve said before, I give less and less credence to the AI benchmarks the labs use as training targets.

Aaron Levie, head of cloud company Box, stated that in their internal tests, GPT 5.5 saw a 10 percentage point jump on accuracy of enterprise content tasks vs. GPT-5.4. They test AI models on a variety of tasks that “represent real work scenarios in financial services, healthcare, public sector, and more industries, dealing with financial filings, clinical records, policy documents, and creative content.”

Levie stated that, “Overall GPT 5.5 is much better at advanced reasoning tasks, better at data analysis, handling complex context, and more. It will be a big jump for enterprise knowledge work agents.” That’s the kind of independent, outside evaluation I trust much more than the benchmark evals in the labs’ own press releases.

In my personal use, while I was super-impressed by the new image model, I have not noticed much improvement in the dot-release of the language model. I don’t do those complex enterprise tasks that Box studied, and even if there were also a 10% improvement on my tasks (e.g., Deep Research about a question) this may be hard to subjectively recognize.

DeepSeek 4

In the United States, DeepSeek released its version 4 the same evening as OpenAI released GPT 5.5, though in China it was the next morning. This was a more impressive upgrade, though that’s mainly because a full year had passed since the release of DeepSeek 3’s thinking model in January 2025.

DeepSeek’s signature whales celebrate DeepSeek version 4 by attempting underwater typography. (GPT-Images-2)

DeepSeek 4 offers many technical advances, such as 1.6 trillion parameters and a one-million token context window. Such strength is unheard of in an open model, and was only reached recently in the frontier closed models in the United States. V4 is also claimed to be extremely efficient in its use of compute, which is important anywhere in the world, given the current compute famine that’s expected to last beyond 2030, but especially important in China which is still catching up on manufacturing advanced AI chips domestically and is subject to a US/NATO blockade on importing these chips.

Of course, tech specs don’t matter to users. They only care what AI can do for them. I put DeepSeek 4 to the test of summarizing an extensive set of research papers. I already had a 3,000 word summary from Gemini Deep Research, but I needed something shorter and easier to read, with less reliance on technical terms. DeepSeek 4 did an excellent job, and in fact produced a better-written condensation of the research than Claude Opus 4.7, which has been my go-to model for writing. Impressive that a Chinese model writes better English than Claude.

One small test doesn’t decide the competition. Here’s an overview of how AI influencers have compared DeepSeek 4 and GPT 5.5 after their release:

(GPT-Images-2)

One of the great abilities of GPT-Images-2 is that we can request it to redraw an entire comic strip in a different style and it will do so, while retaining character consistency.

Here are sample pages from two other cartooning styles I attempted:

I prefer this last cartooning style with hand-lettering. Unfortunately, it’s harder to read than the typeset speech bubbles in the full comic strip I chose to run in this newsletter, especially at the smaller size many subscribers will see in their email programs. (You can always go to the permanent version of any of my articles on www.uxtigers.com and click on any visual to expand it. But of course users should not need to do this as a part of normal reading.) While I am a huge fan of GPT-Images-2, it does tend to use too much text at too small font sizes.

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,187 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: Employer encouragement predicts workplace AI adoption roughly 4x more strongly than training or tool provision alone. This 10-step playbook shows design leaders how to build a culture where AI adoption happens naturally, without mandates, surveillance, or wasted training budgets, while avoiding the cybersecurity pitfalls of “Shadow AI.”

The research is unequivocal: employer encouragement is the strongest predictor of worker AI adoption, far exceeding the effect of training programs or even tool provision. In a recent study by the National Bureau of Economic Research, 47% of workers who received no AI training adopted AI when their employer encouraged it, versus just 10% when their employer did not. That almost 5x difference is a cultural signal doing the work that most managers mistakenly think training should do.

A surprising finding from the NBER research was that formal AI training did not predict higher adoption. Read my extensive analysis of this important study of real-world AI use.

The implication cuts against prevailing HR orthodoxy. If you are a design leader who has been waiting for the “right” curriculum, vendor partnership, or certification program before green-lighting AI use on your team, you are losing months of compounding practical experience that your competitors are gaining right now.

The 10 steps I recommend for building an AI-positive culture in your company. (All images in this article made with GPT-Images-2.)

If training isn’t the primary driver, what is? Building an organically AI-positive culture. Here is my 10-step actionable plan for embedding AI safely and strategically into your organization.

Step 1: Permission, Not Training

The most important thing management can do is remove ambiguity about whether AI use is welcome. In practical terms, this means making an explicit, public, and repeated statement to your team that using generative AI in their design work is not merely tolerated but actively encouraged.

Don’t assume people already know this. Many designers operate under genuine uncertainty: Is AI use “allowed”? Will it be perceived as cheating or laziness? Will using AI-generated outputs reflect poorly on their skills at review time? In the absence of clear signals, your silence is interpreted as discouragement.

There is also a severe and often-ignored consequence to managerial silence: the rise of “Shadow AI.” When leadership fails to explicitly sanction and provide secure AI tools, employees do not simply revert to manual labor to hit their productivity targets. Instead, many turn to unauthorized, consumer-grade models on their personal devices, covertly pasting highly sensitive corporate data or client IP into unvetted systems. Providing clear permission is not just about driving adoption metrics. It is a cybersecurity imperative to prevent data leakage.

Concrete actions:

• Deliver the message personally, not through a company-wide policy memo that people may or may not read. Say it in a team meeting. Say it in your one-on-ones. Post it in the Slack channel. Then say it again three months later. Cultural signals require repetition to take hold.

• Name specific use cases you endorse: “I encourage you to use Claude or ChatGPT to draft your user research syntheses, rewrite feedback for clarity, or explore design directions.” (While also encouraging exploration of new use cases.)

• Name the sanctioned tools explicitly. Ambiguity about which tools are approved drives employees back toward unvetted consumer products.

Ambiguity breeds anxiety. Clear permission breeds innovation.

Step 2: Lead by Example

Permission-granting only works when it is credible. If you tell your team to use AI but they never see you using it, they will read your encouragement as performative. Your visible AI use is the single most powerful demonstration that adoption is safe and valued.

This does not mean announcing “I used AI for this” after every artifact. It means naturally weaving AI into your observable workflow. Share a prompt that worked. Demo a useful output in a team meeting. When you open a document, occasionally let people see you pull it into an AI tool to summarize or critique. If you delegate something, mention that the first draft came from an AI and you iterated from there.

Leaders who refuse to use the tools themselves while exhorting their teams to adopt them are engaged in a form of hypocrisy that is immediately visible to subordinates. The fastest way to accelerate cultural adoption is for the most senior person in the room to become a visibly competent, unselfconscious AI user.

True cultural transformation happens top-down in visibility, but bottom-up in application. You need both.

Step 3: Provide Tools Before You Provide Curriculum

Once you’ve established encouragement, the next priority is access. The research shows that tool provision is the second strongest predictor of adoption, adding meaningful lift even after controlling for encouragement. Training, by contrast, shows no additional predictive power once encouragement and tools are accounted for.

Your budget allocation should reflect this evidence. Before investing in workshops, courses, or AI bootcamps, make sure every member of your team has paid subscriptions to the AI tools most relevant to their work. For most design teams today, that means at minimum a “Pro”-level Big-3 frontier model subscription for text-based tasks, plus access to relevant image-generation and prototyping tools appropriate to your domain.

(Big-3 = Google, OpenAI, Anthropic. This may become Big-5 later this year if xAI and Meta catch up. It does not really matter which top model your employees choose, and there may be benefits to having different staff gain fluency with different vendors, since each has distinct strengths.)

All employees should have a high-tier subscription to at least one provider, such as Google “Ultra” (US $250/month), OpenAI “Pro” ($200/month), or Claude “Max” ($200/month). The cheaper $20/month plans are fine for personal use but not for professional use in an organization serious about AI adoption. Rate limits and model access on the lower tiers will silently cap your team’s productivity in ways you will never see in a dashboard.

Eliminate procurement friction. Do not make people request access through a three-week process that requires a business justification form. The friction of obtaining tools is itself a powerful signal about how much the organization values their use. If an employee can expense a $40 book without approval but needs to submit a ticket for a $20 AI subscription, you have communicated something about organizational priorities, whether you intended to or not.

Activate what you have already bought. Where enterprise-grade tools are available, such as Microsoft Copilot integrated into existing workflows, or Figma’s AI features, ensure they are activated, licensed to the right people, and configured for your team. Many organizations have purchased enterprise AI licenses that sit unused because no one ensured they were deployed to the people who would actually benefit from them.

A rough budgeting guideline. A company that aims to be AI-Forward should allocate at least 15% of each employee’s salary toward that person’s AI tool stack (for non-developer staff), and 33% for developer staff. For an AI-First company, double these percentages; for AI-Native companies, triple them. You don’t need to be tokenmaxxing, but if senior developers in particular cannot spend on AI with abandon, they are not using AI to its fullest potential, and the opportunity cost of constrained usage almost always dwarfs the marginal tool spend.

(The reason I suggest higher token budgets for developers is not that programming is more important than design or user research. It’s simply the case that right now, AI tools for software development are far ahead of AI’s capabilities for other business workflows. And AI developer tools are hungry for tokens. In the future, as AI’s “jagged” frontier keeps shifting, more powerful, more hungry, AI capabilities will likely become available for other functions, and token budgets should be adjusted up accordingly.)

Step 4: Create Space for Experimentation Without Mandating Specific Workflows

Encouragement and tool access set the conditions for adoption, but the specific ways designers integrate AI into their work should emerge organically from practice, not from top-down workflow mandates. The research shows that AI users apply the technology across an average of four different task types, with the most common being writing communications, searching for information, interpreting and summarizing content, and generating new ideas. This diversity suggests there is no single “right way” to use AI in design work, and trying to impose one will suppress the creative adaptation you need.

New applications will also emerge every few months as more capable models are released, unlocking possibilities previously beyond AI’s reach without prohibitively high hallucination rates. A workflow you mandated in Q1 may be obsolete by Q3.

Create structured experimentation time. Dedicate explicit time in your team’s schedule (perhaps an hour per week or a day per month) to exploring AI tools in the context of current projects. Make this time legitimate and protected, not something people squeeze in between deadlines. If it isn’t on the calendar, it isn’t real.

Add a lightweight sharing mechanism. A standing 5-minute agenda item in your regular team meeting, where one person shares an AI experiment (what they tried, what worked, what did not), normalizes the practice and spreads practical knowledge far more effectively than formal training. Rotate the presenter so that every team member eventually has to articulate something they have learned. The act of teaching compounds the learning.

Insist that experiments use real work. A designer who discovers that AI dramatically speeds up their competitive audit process will adopt that practice permanently. A designer who completes a generic prompt-engineering exercise will likely forget it within a week. Abstract training evaporates; lived experience sticks.

Mandating a rigid “Standard Operating Procedure” for AI is a fool’s errand because the technology moves faster than corporate governance. For that same reason, hiring even good strategy consultants to bring you the best practices from their other clients won’t work. Those practices may indeed have been “best” last quarter, but are second-best now. You must discover what works in your company by experimentation, because nobody knows.

Localized, tacit knowledge is your most valuable asset in the AI transition.

Step 5: Address Quality and Ethics Directly, Not Through Prohibition

Many design managers hesitate to encourage AI use because they worry about quality control, intellectual property, confidentiality, and ethical concerns. These are legitimate. But addressing them through prohibition or silence is worse than addressing them through clear, practical guidelines.

Establish explicit guardrails in plain language. Specify what types of information should never be entered into external AI tools: client confidential data, personally identifiable information, proprietary research findings, unreleased product specifications, whatever applies in your context. Specify review expectations for AI-assisted outputs: all AI-generated copy must be reviewed by a content specialist before inclusion in deliverables; AI-generated visual assets must be evaluated against brand guidelines by a visual designer before use; AI-drafted research summaries must be verified against the source materials before being shared with stakeholders.

Treat AI like any other professional tool. You do not prohibit designers from using stock photography because some of it is bad. You establish quality standards and trust professionals to meet them. Apply the same logic to AI. Define the boundaries clearly, then trust your team to operate within them.

Be honest that ethics are evolving. Tell your team directly that reasonable people disagree about some questions (training data provenance, attribution, disclosure to clients) and that the standards are still being worked out across the industry. Encourage critical thinking about when AI use is appropriate and when it is not, rather than issuing blanket rules that will become outdated as the technology and its social context evolve.

Ethical and security boundaries provide freedom. When employees know the red lines, they can innovate fearlessly within the safe zones.

Step 6: Address Job Security Anxiety Head-On

Underneath most resistance to AI adoption lies a question employees rarely ask out loud: Am I training my own replacement? If you do not address this directly, it will quietly throttle every other step in this plan.

Silence on the question is read as confirmation of the worst case. Vague reassurances (“AI will augment, not replace”) are read as corporate-speak. What works is specificity.

Be honest about what you can and cannot promise. You probably cannot guarantee that headcount will never change. What you can commit to is that people who develop strong AI skills will be more valuable, not less, and that the team’s mandate is to do better work (and likely more work), not to produce the same work with fewer people. If layoffs are genuinely on the table, say so; pretending otherwise destroys trust far more than the layoff itself would.

Redirect time savings toward visibly valuable work. If AI saves a researcher five hours a week, and those five hours are invisibly absorbed into more Slack or more meetings, adoption will stall. If they are visibly redirected into deeper user interviews, a new project, or strategic work that raises the team’s profile, adoption compounds. People adopt tools that make their jobs better. They resist tools that make their jobs disappear.

Create a “skills runway.” Explicitly commit team budget and time to helping everyone develop the next layer of skills that AI does not yet do well: systems thinking, stakeholder facilitation, strategic framing, ethical judgment. Framing AI adoption as part of a career-development arc is far more motivating than framing it as a productivity mandate.

Step 7: Reward Outcomes, Not Methods

The research linking AI adoption to performance-oriented management practices is striking. Countries and firms where performance is rewarded, promotions are based on merit, and poor performance is addressed also have dramatically higher AI adoption rates. The mechanism appears to be that performance-focused cultures create natural incentives to seek out productivity-enhancing tools: people who have skin in the game of their own output find AI on their own.

For design managers, this means evaluating your team on the quality, speed, and impact of their work, not on the specific methods they used to produce it. If a usability specialist delivers an excellent research synthesis in half the usual time because he or she used AI to process interview transcripts, that is a success to be recognized, not an asterisk to be noted.

Conversely, be alert to the possibility that some team members resist AI not because they have thoughtfully evaluated its limitations but because adopting new tools feels threatening to their professional identity. The most common reason workers give for not using AI is that they believe it cannot help with their job. A belief that, given the breadth of current AI capabilities, is already incorrect for most knowledge workers, and will be even more false soon. A performance-oriented culture naturally surfaces this disconnect: when peers using AI are producing better work faster, the incentive to reconsider one’s assumptions becomes tangible.

Guard against performative AI use. The goal is better design outcomes, not more prompts entered. Some tasks genuinely do not benefit from AI, and experienced designers should be trusted to make that judgment. If you start celebrating AI use itself rather than the work it produces, you will get a lot of theater and very little improvement. (Again, don’t tokenmaxx; outcomemaxx.)

Step 8: Account for Uneven Adoption Within Your Team

The data shows AI adoption varies significantly by age, education, and individual disposition. Within any team, you will likely have enthusiastic early adopters, cautious experimenters, and outright skeptics. A one-size-fits-all approach will either bore the first group or overwhelm the third.

Leverage early adopters as internal champions. Pair them with more hesitant colleagues on specific projects where AI might add value, and let practical demonstration do the persuading that no amount of managerial exhortation can achieve. Peer influence is powerful. Seeing a respected colleague use AI effectively to solve a real problem is more persuasive than any keynote about AI’s theoretical capabilities.

Create optional “AI office hours” staffed by your most fluent users, where colleagues can bring specific problems (“I have 200 pages of interview notes, what should I do?”) and leave with working solutions. This outperforms centralized training because it is demand-driven, specific, and immediately useful.

Respect persistent skepticism, but set clear expectations. For team members who remain skeptical after genuine exposure, respect their professional judgment while continuing to hold them to the same performance standards as everyone else. Some resistance will dissolve naturally as tools improve and use cases become obvious. Some may reflect legitimate concerns about quality, reliability, or appropriateness that you should listen to carefully. Your skeptics may be seeing real problems your enthusiasts are overlooking.

Step 9: Track Adoption and Impact, Even Informally

You cannot manage what you do not measure. But you also should not create surveillance mechanisms that undermine the trust you are trying to build. A reasonable middle ground is to incorporate AI-related questions into your existing team rituals.

In project retrospectives, ask whether AI tools were used, where they helped, and where they fell short. Always assume that current AI is imperfect: that is fine, and not to be punished. But do learn from the mistakes, and remember that what failed this year could very well succeed next year as models improve.

In quarterly planning, ask team members to identify one or two areas where they would like to experiment with AI in the coming quarter.

In performance conversations, discuss how each employee’s toolkit has evolved and whether they are staying current with available technologies.

Over time, you should be able to observe whether your team’s AI adoption is increasing, whether the tools are generating real resource savings or quality improvements, and whether certain use cases are proving more valuable than others. Feed this information back into your tool-provisioning decisions, your team’s knowledge-sharing, and your own understanding of where AI adds genuine value in the design process versus where it remains more hype than substance.

A simple dashboard of three numbers, monthly AI tool spend per head, number of experiments shared in team meetings per quarter, and estimated hours saved (self-reported), is more than sufficient for most teams. Resist the temptation to build elaborate measurement infrastructure. It is almost always a substitute for doing the actual cultural work.

Step 10: Connect AI Adoption to Strategic Design Goals

Finally, frame AI adoption not as a technology initiative but as a means to better design outcomes and higher business profitability. The research shows that AI users already report meaningful time savings: about 5.8% of work hours on average. (This will grow substantially every year for the next decade.) For a design team, that recovered time can be redirected toward activities AI currently handles poorly, such as deep user empathy work, complex systems thinking, cross-functional alignment, and the kind of creative synthesis that emerges from sustained human attention to ambiguous problems.

The strategic case for AI in design is not that AI will do the design work for you. It is that AI can absorb a meaningful share of the routine cognitive labor (drafting communications, summarizing research, generating initial variations, searching for precedents) that currently competes for the same hours you would rather spend on higher-order design thinking.

Position AI adoption within this frame for your team, and you accomplish two things simultaneously. You give people a compelling reason to experiment with the technology, and you reassure them that their human skills remain central to their value. Both messages are important, and both are true.

Redesign the workflow, don’t just automate the task.

Conclusion

The companies that will win the next decade are not the ones with the most elaborate AI training programs or the largest enterprise licenses. They are the ones whose leaders understand that AI adoption is, fundamentally, a cultural problem dressed up as a technology problem.

The NBER data is clear:

• Permission beats pedagogy,

• Tools beat training, and

• Culture beats curriculum.

Build a positive AI culture in your company.

My 10 steps are not a rigid sequence; they are a reinforcing system. Permission without tools is empty. Tools without permission are unused. Tools and permission without outcome-based rewards produce performative adoption. Outcome-based rewards without attention to job security produce quiet sabotage.

If you are a design leader wondering where to start, start with steps 1 and 2 this week. Tell your team, out loud and in writing, that you want them using AI. Then use it yourself, visibly, tomorrow morning. Everything else in this plan becomes easier once those two signals are in the air.

The cost of getting this wrong is not that your team will be a little less productive. It is that your best people will safeguard their career future by leaving for companies where their AI fluency is welcomed, while the colleagues who stay will be doing their AI work on unauthorized personal accounts, pasting your client data into systems you cannot audit. The choice is not whether your team uses AI. It is whether they use it in the open, with your support and guardrails, or in the shadows, without them.

Summary of the 10 steps I recommend in this article.

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,153 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: Translating social media posts by default makes online communities international | Indian film studios embrace AI | AI summaries rescue users from context decay | AI agents now account for almost as much website traffic as human search | Software customers demand agent-first design | “You Are Not Talking to Somebody Who Woke Up a Loser”

UX Roundup for April 20, 2026 (Nano Banana 2)

AI Facilitates International Understanding and Community

The social media platform X (previously known as Twitter) launched a new feature a few weeks ago: automatic translation of posts from the author’s language to the user’s language. For example, since I am logged in from the United States, I see all posts in English, even when they were originally written in another language, such as Japanese or Arabic. (You can set preference options to retain the original wording for posts in additional languages you read, such as Danish, in my case.)

The translations are done in the background by Grok (another xAI product) and are excellent. The English versions of posts are always fluent and eloquent, and I have not been able to distinguish them from those of native-language speakers. If anything, posts by Americans and Brits contain more typos and awkward phrasings than Grok’s translation of foreign-language posts.

The translations are highly accurate, at least when I have compared a translated post with the original in a language I read. (X offers a one-click option on all translated posts to reveal the original text.)

A possible next step may be to not just translate the original text but improve it. However, I don’t recommend that. (Nano Banana Pro)

Translations are not just from other languages into English. The translations work between any pair of languages, so that, for example, a user in Korea could post in Korean and a user in Brazil would see that post in Portuguese.

It has long been possible to paste any text into an AI-powered translator, whether via the standard chatbot UI or a specialized translation service. What’s different with the new X translations is that they have already been automatically applied to all posts before you see them, meaning any user’s social feed is truly international by default. That little extra step required in the past was enough to prevent most users from translating most posts, and in fact, most social media platforms would hide foreign-language posts by default.

Auto-translations place users in all countries on an equal footing, whereas in the past, social media provided a substantial advantage to users in English-speaking countries. First, writing in your native language is always easier, even for people with foreign-language skills. Being able to do so encourages more people from more countries to post. Second, people almost always express themselves better in their native language than in a foreign language. This means that in the past, posts from foreigners felt less polished than posts from native-language speakers. This again made readers feel, subconsciously, that foreigners were stupid because they didn’t write as well.

Of course, translation doesn’t make a stupid person smart, but it does make all users, from all countries, appear as smart as they actually are, without the degraded communication abilities that come from trying to express oneself in a foreign language.

Will auto-translation on X bring world peace? No. But it does increase international understanding and makes the social media platform more interesting. Score one for AI making a positive contribution to society.

Translation by default = more international communication, for a truly worldwide community on social media. (Nano Banana Pro)

Bollywood Increasingly Using AI For Its Films

Continuing the topic from the previous news item of how AI translation broadens information access: India has the world’s largest film industry, and Bollywood studios are now using AI to translate, dub, and lip-sync movies into multiple languages.

Bollywood is going all in with AI. (Nano Banana Pro)

India has around 22 languages with more than 5 million native speakers (a number I picked because it’s the number of native speakers of my native language, Danish). Yes, Hindi is much the biggest with 528 M native speakers, but whether from a business perspective or a cultural perspective, you don’t want your films to miss out on ticket sales to people who use Bengali (97 M speakers), Marathi (83 M), Telugu (81 M), and so forth.

Film the actress once, and she will be a native speaker of 22 different languages in the final movie. (Nano Banana Pro)

Reuters recently had an interesting article about the use of AI in the Indian film industry. Converting films from monolingual to multilingual virtually for free stood out to me. However, AI is used much more widely in India than in the American film industry, which is chained down by ridiculous union rules. The article quotes industry insiders who estimate that current AI can slash production costs to 20% of traditional budgets and cut timelines by a quarter. (And we absolutely expect the next generation of video AI tools to do better.)

AI is particularly popular for mythological and religious films, which makes sense given the appearance of many important Hindu gods.

Films based on mythological stories are extremely popular in India, and are immensely cheaper to make with AI. (Nano Banana Pro)

Naturally, early AI adoption suffers from teething problems. JioStar recently released an AI-generated adaptation of the epic Mahabharat. Its reception? A dismal 1.4 out of 10 on IMDb, with users complaining about lip-sync issues and unnatural styling.

Reach for the future. Don’t get left behind with yesterday’s tools. This applies to anybody, even if you’re not making movies. (Nano Banana Pro)

Right now, Indian AI films may still not be as good as the (much more expensive) films made the old-fashioned way, but we should remember that what’s now considered old-school filmmaking was once innovative and suffered its own quality problems. What’s more interesting is to consider that using our new capabilities for full-scale creative projects is the way to grow a community of skilled creators who understand the new media form. These people will be better positioned to create truly great work in a few years, as AI improves, than latecomers, who won’t start gaining creative experience until it’s too late to catch up.

You become good at using AI by using AI. You become creative at using AI by creating with AI. Countries with fewer inhibitions against innovation will dominate the future. (Nano Banana Pro)

Reuters quotes Dominic Lees (a British film school professor) as saying, “If they can deliver, then the shift in AI filmmaking will be to India.” I think he may just be right.

The future belongs to those who show up. In creative filmmaking, this may be India. (Nano Banana Pro)

Whatever else happens, audiences will get more, better, and cheaper entertainment as the use of AI in creative industries expands. (Nano Banana Pro)

Context Reboarding: How AI Summaries Rescue Us from Context Decay

Work fragmentation is the curse of the modern knowledge worker. You are deep in a complex task, an urgent email demands your attention, and you step away. When you return, you stare blankly at your screen, suffering from cognitive discontinuity. You wonder, “What on earth was I trying to do?”

Any interruption to a focused task causes a substantial disruption to the task flow when the user tries to resume the task. (Nano Banana 2)

To successfully resume an interrupted task, users must rebuild two distinct mental structures:

1. Working context: The physical artifacts. Which files were open? What documentation was I reading?

2. Mental context: The overarching goal. What was my intent? Why was I looking at these specific artifacts?

For decades, we have forced users to either write meticulous manual notes or waste precious cognitive cycles rereading their own work to reconstruct their mental model. Both are severe usability anti-patterns, relying on human discipline and perfect memory, neither of which exist in the real world. But recent advances in AI offer a superior interaction paradigm.

A fascinating new paper by Alexander Lill and colleagues from the University of Zurich in Switzerland, TaCoS: Generated Context Summaries for Task Resumption, brings rigorous empirical data to this exact problem, proving that AI can bridge the cognitive gap.

The Experiment: Testing Resumption Cues

The researchers built an IDE (Integrated Development Environment) extension that quietly monitors a software developer’s activity: code edits, file navigation, and browser history. When the developer returns from an interruption, the AI generates a Task Context Summary, featuring the inferred user intent and a structured list of recent actions.

To test this, the researchers ran a controlled lab experiment with 32 software developers. They interrupted the programmers mid-task and forced them to resume 1 to 7 days later, ensuring total context decay. Each participant tested three resumption cues:

• Manual Notes: The traditional method, where users typed their own reminder before leaving.

• Timeline: A chronological, exhaustive visual history of recent clicks and file edits.

• Generated Summary (TaCoS): The automated AI synthesis of the user’s overarching intent and completed steps.

The three study conditions for helping users reestablish their context after an interruption: the user’s own notes, a full history of what the user had been doing, and an AI-generated summary. (Nano Banana 2)

Empirical Findings: Synthesis Beats Raw Data

The results were a resounding victory for automated summarization.

The AI-generated summary produced the shortest resumption lag (the time it takes a user to make their first navigation action), dropping by 35% from 69 seconds for notes to 45 seconds for AI. AI-generated resumption summaries were also overwhelmingly preferred, with 27 out of 32 users choosing the AI summary over their own handwritten notes.

Users strongly preferred AI-generated summaries to their own notes, and they performed better when relying on AI. However, these users were all nerds, so it remains to be seen whether the findings will generalize to other knowledge workers. I suspect that the performance edge for AI will remain, but that the subjective preference may not be as strong. (NotebookLM)

Decent, but less impressive, results were recorded for edit lag (the time to resume actual work), which was 186 seconds in the manual condition and 172 secs when using AI-generated summaries, for an 8% improvement. The number of successfully completed subtasks was 1.25 in the manual condition and 1.35 for AI-generated summaries, which was also 8% better.

Why was AI better than the users’ own notes? Because human beings are notoriously unreliable at documenting their own mental state during an unexpected interruption. The AI summary provided the “why” (inferred intent) alongside the “what” (clickable links to relevant files) with zero manual effort.

People don’t know what they will need later, so their notes are not as useful as those automatically generated by AI. (Nano Banana 2)

Interestingly, while the raw timeline of edits technically produced a slightly higher task success rate (1.56), users complained it was visually overwhelming. As I have advised for decades, presenting raw system logs is not a user interface. Users need synthesized meaning, not a data dump. Assuming that giving the user all the data is the same as giving them the right data is a classic usability failure.

The pure data dump of everything the computer knew was overwhelming. (Nano Banana 2)

However, the study revealed one critical limitation of current AI: while the LLM was excellent at retrospective summarization (what happened), it failed at prospective planning (what to do next). The self-authored manual notes almost always contained the immediate next step. The developers relied on the AI to remember the past, but they needed their own brains to chart the future.

Generalizing to Slow AI

These findings generalize far beyond software development; they provide empirical validation for the necessity of Context Reboarding, a concept I introduced in my recent article on Slow AI.

As AI agents begin executing complex tasks that take hours or days to complete, we are effectively reviving the batch-processing paradigm of the 1960s. Frequent turn-taking vanishes. When a 20-hour AI task finally finishes, your mental context has completely decayed. You cannot be expected to remember the exact nuances of a prompt you wrote yesterday.

If your product features long-running AI tasks, you must design a Return Recap (or Resumption Summary). You cannot simply flash a generic “Task Complete” toast notification. The UI must explicitly welcome the user back by stating:

• Original Intent: This is what you asked me to do.

• Conceptual Breadcrumbs: Here are the key decisions I made while you were gone.

• Current Status: Here is the outcome and the exact artifacts produced.

The Future of AI UX: Hybrid Interfaces

The ultimate UX solution for the future of AI is a hybrid design: an AI-generated summary of past actions and intent, paired with a lightweight, user-authored “next step” input field. Let the computer do the tedious work of tracking history, while empowering the human to do the high-level strategic planning.

We must stop pretending that users possess infinite working memory. The TaCoS experiment proves that automatic, intent-driven summaries drastically reduce the cognitive friction of task resumption. Whether a user is interrupted by a ringing phone or by a 24-hour AI processing cycle, the UX mandate remains the same: do not make the user remember. Place the memory burden on computers. They are superior to our poor brains at keeping information from decaying over time.

AI Agents = 15% of Website Traffic

The way people find information online is increasingly mediated by AI agents rather than direct human queries. New data from SEO vendor BrightEdge shows how profound this shift has become. According to the company, AI agent requests, largely coming from systems like ChatGPT, Perplexity, Gemini, and Claude, have reached 88 percent of human organic search volume. These agents already generate about 15 percent of total website traffic as of April 2026. If growth continues, AI agent requests will surpass human search activity later this year.

Why does this matter? AI agents are not merely gathering information; they are making decisions on behalf of users. BrightEdge warns that at key decision‑making moments, such as choosing a product or service, the content that agents retrieve and interpret may determine which brands succeed. Yet most companies are unprepared. Only 19 percent have specific directives for AI bots, while 81 percent treat them like ordinary crawlers. Traditional robots.txt files may inadvertently block AI training agents or allow retrieval agents to misrepresent content, leading to missed opportunities or misaligned brand messaging.

AI accounts for an ever-growing share of website traffic, meaning that content strategists need to adapt to the new GEO realities. However, don’t forget the human users for now. (Nano Banana 2)

It’s time to develop AI agent strategies. This involves deciding which agents to welcome, what content they may access, and how to structure data for retrieval and training. From a UX perspective, designers should consider how AI agents summarize or rephrase content. Additionally, transparency about AI agent traffic is essential; companies need analytics tools that distinguish between human and agent visits to understand how their content is being used.

The findings also highlight the growing importance of Generative Engine Optimization (GEO), the practice of optimizing content for conversational AI. Unlike traditional SEO, GEO prioritizes concise, unambiguous facts and context that AI can easily parse. For designers and content strategists, this means rethinking information architecture, ensuring that key facts are prominent and that language is clear. Failure to adapt could mean that AI agents will simply ignore a company’s content, leaving users with incomplete or incorrect information.

BrightEdge’s report signals a shift in digital discovery. UX professionals must recognize that their audience is shifting from humans to AI agents. During this transition period, when humans still visit websites, designing websites and content that serve both human visitors and AI agents is critical for maintaining visibility and trust in an AI‑mediated world.

Headless Software = Agent-First Design = No UI

As I have been saying for a year, we’re heading for a world without user interfaces in the traditional sense. It is likely that humans will only interact with their AI agents, and the agents will then handle interactions with online services (formerly known as websites) and applications on behalf of the users. The previous news item, showing that AI agents already almost equal traditional search traffic to websites, demonstrates that this trend is well underway.

The AI handles all the interactions that previously required separate user interfaces for each website. (Nano Banana 2)

Yes, the AI agents will need UX design, because they do have a user interface, so it’s not literally true that there will be zero UI. However, AX > UX, to the extent that the agent experience becomes vastly more important than the little UX that remains.

Agents-First is becoming the design goal for many new projects. (Nano Banana 2)

Aaron Levie (head of cloud storage provider Box) wrote an interesting article about trends in AI agent use, summarizing meetings he has had with IT leaders in large enterprises across banking, media, retail, healthcare, consulting, tech, and sports, to discuss agents in the enterprise. Note that these are very conservative industries. (I have found Levie to be one of the most insightful commentators on where AI and UX are going, probably because — besides clearly being very smart — he has the finger on the pulse as both leading an AI-First company and meeting frequently with a wide range of people trying to implement AI in their companies. This guy doesn’t just sit in his office.)

Levie reported that “Headless software dominated my conversations. Enterprises need to be able to ensure all of their software works across any set of agents they choose. They will kick out vendors that don’t make this technically or economically easy.” (“Headless” software runs without a built‑in graphical user interface and typically exposes its capabilities via APIs or other non-visual interfaces, so any number of separate “heads” [frontends] can sit on top of it.)

Computer geeks are notoriously bad at naming, as we’ve seen with AI model names ever since ChatGPT. (Nano Banana 2)

Being “headless” means that software can be fully controlled without a screen (the “head,” or terminal in old-school IT systems).

It’s not simply the technical ability to use software without a GUI. That API interface must be easy for the agents to use: AX!

Companies are starting to reject software vendors that don’t make their software “headless,” so that it can be easily operated by an AI agent. (Nano Banana 2)

What’s the balance between UX and AX? We will soon get to the point of Agent-First Design, similar to the “Mobile-First Design” that was all the rage 10 years ago. For now, this doesn’t mean “no humans,” any more than the mobile-first era meant websites could ignore desktop visitors. It’s more the question of what’s more important and where you lavish the most resources, as the balance gradually shifts. It’s probably at least 10 years until we get to a true “No UI” situation, where companies only need to design for AI agents to access their software and websites.

Think normal people enjoy using websites? Think again! Most people will be happy if they never have to visit an online banking site ever again because their agent deals with the bank on their behalf. (Nano Banana 2)

While many people have wised up over the last year to the need to design for agents, Levie made a second observation in his article that I don’t see mentioned nearly as often: “Despite Silicon Valley’s sense that AI has made hard things easy, the most powerful ways to use agents is more “technical” than prior eras of software. Skills, MCP, CLIs, etc., may be simple concepts for tech, but in the real world, these are all esoteric concepts that will require technical people to help bring to life in the enterprise.”

A call for more usability, if I ever heard one. (Levie thinks that the solution is to have more technical people to explain AI to the users. I think the solution is to make AI agents easier.)

AI agents are currently too difficult for many regular businesspeople to set up. We need better agent usability. (Nano Banana 2)

Profitable business use of AI requires new workflows designed for this technology, which again means that domain experts across the business must feel comfortable setting up the agents.

We’re not quite at the point yet where Agent Experience (AX) actually is more important than UX every time, but the balance is tipping, so this scenario will be true in more companies every year. (Nano Banana 2)

Suno Sale

Suno is currently the best AI model for making songs. My latest creation: Hamlet, the Music Video (YouTube, 4 min.), which features really neat animations made with Seedance 2.0.

Suno is offering 20% off full-year subscription plans, with the sale ending tomorrow, Tuesday, April 21, at 9 PM USA Eastern Time.

(This is not a sponsored post: I am a satisfied, paying customer, so I’m the one giving Suno money. They don’t give me anything.)

However, despite liking Suno, I recommend against this offer unless you are a huge music creator. Annual plans should be avoided for most AI services because the likelihood is low that any given service will remain the best for a full year. This is less of a problem for the Big-3 foundation models (Google, OpenAI, and Anthropic), as well as the two contenders that may make it to a potential Big-5 list later this year: xAI and Meta. Even though the big AI models also alternate every few months on who’s ahead, the difference remains marginal. Let’s say that OpenAI’s rumored “Spud” model takes the lead in April. That lead may only last a few weeks until Google’s rumored release in May. Unless you have a use case where the last few percent in AI capability will make or break your product, then just pick one model and stick with it.

However, for more narrowly targeted AI models, such as music, image, or video generation, the differences can be larger and not close as quickly after one model pulls ahead of the pack. You don’t want to be stuck having prepaid for the laggard for another half year.

My advice on AI subscriptions. (Nano Banana Pro)

Gemini 3.1 Pro Time Horizon = 6 Hours, 24 Minutes

METR has released its measurements of the T50 time horizon for Google’s Gemini 3.1 Pro with thinking level “high.” This model scores a time horizon of 6 hours 24 minutes, which is 12% better than GPT 5.4 (running at xhigh reasoning) which has a time horizon of 5 hours and 42 minutes.

Claude Opus 4.6 still holds the gold medal as the frontier with the longest T50 time horizon, of 11 hours 59 minutes, which is a whopping 87% better than Gemini 3.1 Pro. When you look at the chart at the MERT page linked above, you can see how Opus 4.6 is a true outlier in the current Big-3 model race toward superintelligence.

Of course, both OpenAI and Google are likely to have a comeback with their next models.

Gemini 3.0 Pro has a T50 of 3 hours 44 minutes, meaning that the leap from 3.0 to 3.1 was 71%. An immense gain for Google in only 3 months. (3.0 was released November 2025 and 3.1 was a February 2026 release.)

Google is heavily rumored in Silicon Valley to have its next model cooking for a May 2026 release, which would represent another 3 months of AI advances. If Google can bag another 71% improvement, they will virtually close the gap with Opus 4.6.

Google gains, but Claude is still King of time horizons, which is why most developers use it for their long-running agentic coding tasks. (Nano Banana 2)

Google actually needs to beat Opus 4.7, not 4.6. Unfortunately, we don’t know yet how good this model (launched April 16) is on the METR time horizon metric, and it’ll probably take a month or more for METR to complete its research. 4.7 outscores 4.6 on the standard AI industry benchmarks, but those benchmarks have less and less credibility as they get saturated and as AI labs train models specifically to beat the benchmarks.

Opus 4.6 was an outlier, relative to the “standard” exponential improvement of AI models over time for METR time horizons. I have two opposite theories for what this means for the Opus 4.7 score: One possibility is that the “miracle” that bumped 4.6 up so much was a one-off and we’ll see a reversal to the norm with smaller improvements in subsequent versions (such as 4.7), until Claude is back on the trend line. The other possibility is that Anthropic has cracked recursive self-improvement and that the new norm is ever bigger gains. In this second case, the old trendline becomes irrelevant and we’ll need to establish a new trend for the speed of RSI. The trillion-dollar question is obviously whether Anthropic will be the only lab to realize benefits from RSI, or whether they were just the first to release models within this paradigm. The history of AI suggests that the other labs will discover the same magic sauce and start accelerating on that same new trendline. The next half year or so will show whether everybody gets on the same old trendline or whether everybody move to a new trendline. (Or as a distinctly unlike third possibility, that Anthropic accelerates on a new line and the other labs stay on the old line.)

New, faster trendline for improving AI time horizons? Or will everybody revert to the mean predicted by the old trendline? A single data point clearly can’t differentiate between these two scenarios, so we will have to wait to find out. (Nano Banana 2)

As a reminder: the T50 time horizon is the time it would take a human expert to perform a task that the AI can do with 50% accuracy. In this case, Gemini probably executes most of these tasks in less than an hour, but the “time horizon” is measured as the human equivalent.

“You Are Not Talking to Somebody Who Woke Up a Loser”

During a recent appearance on the Dwarkesh Podcast, Nvidia CEO Jensen Huang delivered a brutally honest, meme-worthy statement regarding his competitive drive: “You are not talking to somebody who woke up a loser.”

My Hero: Jensen Huang may have delivered the defining quote of the modern era. I love his optimistic, can-do attitude. (Nano Banana 2)

Every UX professional needs to print this quote and tape it to their monitor. Lately, observing the self-pitying discourse in the design community, it feels like an entire industry woke up feeling exactly that way. You all need to stop thinking as losers about AI and how it will take all the jobs and eliminate UX design.

This pervasive doom-loop is fundamentally backwards. Rather than an apocalyptic threat to our livelihoods, AI is the greatest opportunity for the last two hundred years. The Industrial Revolution produced roughly a 50-fold increase in average living standards across 200 years. AI will compress a comparable transformation into 20 years, perhaps less. When humanity gets dramatically richer, what do rich people demand? Higher quality. Better experiences. Beautifully crafted products. Thoughtfully designed services. The history of affluence is the history of escalating design expectations. Peasants tolerated mud floors; their descendants commissioned Chippendale furniture. Today’s users tolerate text-box interfaces; tomorrow’s will demand AI experiences that make current design look medieval.

Stop with the negative thinking, already. (Nano Banana 2)

AI will make us all unfathomably rich. Just as we already live like kings compared to our ancestors who worked the fields from dawn to dusk. And the rich don’t tolerate friction. They demand elegance, so they will turn away from bad UX. (Nano Banana 2)

And speaking of that text box: have we collectively lost our minds? The dominant interface paradigm of the most transformative technology since electricity is a chat window. A chat window! Usability testing shows current AI tools scoring abysmally against every heuristic that matters. Users cannot discover capabilities. They cannot recover from errors. They have no mental model of what the system can or cannot do. Feedback is inconsistent. Consistency is nonexistent. Visibility of system status is a running joke. This is not a triumph of AI; it is an indictment of AI UX. Someone will redesign these interfaces properly, and that someone will capture enormous value. Why not you? Did you wake up a loser?

The rich possibilities of AI are currently trapped inside a text box. We’re in the extremely early stage of AI interaction design. (Nano Banana 2)

Capturing the vast economic value of AI will not happen automatically. It necessitates immediate action across two massive frontiers, requiring extreme UX investments both in better UI for all these AI tools and in redesigning all enterprise workflows in all companies, using service design methodology to make them optimal for profiting from integrated agentic AI.

Virtually every single enterprise workflow will have to be redesigned for optimal AI integration. (Nano Banana 2)

Let us address the tools first. By relying on prompt engineering, the tech industry has inexplicably retreated to a conversational paradigm that is effectively a dressed-up command-line interface from the 1960s. This severely violates our most fundamental usability heuristic: recognition over recall. Expecting mainstream users to guess hidden system capabilities and articulate perfectly structured text prompts is a catastrophic usability failure. It creates a severe articulation barrier, shifting the cognitive burden from the computer back onto the user. We must move beyond the lazy, blinking cursor to invent rich, discoverable graphical user interfaces that actually guide human-computer interaction.

The dominant UI for AI is a time warp bringing back the bad old days. (Nano Banana 2)

Usability heuristic number 6, recognition over recall, is brutally violated by most current AI user interfaces. (Nano Banana 2)

Second, the true economic payoff extends far beyond individual consumer applications. You cannot simply drop a smart agent into a broken, legacy business process and expect a positive return on investment. The introduction of autonomous capabilities requires us to dismantle and completely rebuild the systematic way corporate work is conducted.

UX practitioners must step up to map out these complex, cross-functional processes from end to end. We are tasked with designing entirely new corporate ecosystems where human workers and autonomous agents operate in tandem. Crafting these intricate hybrid workflows to determine precisely who does what, and how exceptions are gracefully handled by human overseers, will require millions of hours of skilled, rigorous design work to actualize the promised productivity gains.

Stop whining about AI’s downsides, start fixing them. (Nano Banana 2)

The AI revolution is not the end of the UX profession; it is the ultimate full-employment act for designers. We have a world to fix. Stop whining about artificial intelligence, recognize the disastrous usability of today’s AI, and embrace the monumental challenge ahead of us. Stop waking up a loser.

Redesigning the way everything works is the opportunity of a lifetime. (Nano Banana 2)

Jensen Huang made many interesting points during his appearance on the Dwarkesh Podcast. Give the full 103-minute episode a listen. My apologies to Dwarkesh Patel for Nano Banana 2 drawing him without his trademark beard.

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,112 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: AI is saving time at scale, reshaping enterprise behavior, and exposing a new design problem: products must be built for uneven capabilities, uneven adoption, and an uneasy public. While macroeconomic data shows immense productivity gains and US enterprise dominance, micro-level metrics reveal usability flaws, job losses among juniors, and severe algorithmic trust issues.

Two heavyweight reports published in the last few weeks provide data-driven snapshots of how people actually use AI and how it performs in real companies.

The first is an NBER (National Bureau of Economic Research ) working paper, Mind the Gap: AI Adoption in Europe and the U.S. It focuses on workplace adoption: who uses AI on the job, what explains the cross-country differences, and how much time users report saving.

The second is Stanford HAI’s Artificial Intelligence Index Report 2026, a wide-angle survey of technical performance, investment, public sentiment, environmental cost, and governance that clocks in at 423 pages.

The two reports are not tracking one adoption curve but at least three: consumer familiarity, workplace permission, and workflow penetration. Confusing these curves produces many of today’s arguments about whether AI is overhyped or underhyped. A country can look mediocre on population-wide adoption and still be aggressive inside enterprises. A company can report broad experimentation while seeing very little share-of-work penetration. For product strategy, the denominator matters. You are not designing for ‘AI users’ in the abstract. You are designing for people at very different stages of familiarity, authorization, and dependence.

I’ll discuss each report in turn, then give you my takeaway recommendations on what they mean for you.

The NBER Report and the Business Reality

This report addresses a fundamental macroeconomic question: Will AI exacerbate the existing productivity gap between the United States and Europe? To find out, the authors fielded massive, internationally comparable worker and firm surveys in 2025 and 2026.

Their data refutes myths about how AI is entering the workforce and reveals the hidden psychological drivers of enterprise software adoption.

The NBER report focused on high-level measures of AI productivity, and compared these metrics between countries. (GPT Image 1.5)

The US Dominates Enterprise Adoption and Intensity

The data shows a decisive, widening victory for the United States in workplace AI diffusion. In early 2026, 43% of US workers reported using generative AI for their jobs. Compare this to an average of just 32% across the six major European economies surveyed (Germany, France, Italy, Sweden, the Netherlands, and the UK).

Between mid-2025 and early 2026, AI adoption grew in every country surveyed. But it grew fastest in the United States, adding 3.6 percentage points. In the three European countries with the lowest adoption rates (Germany, France, and Italy), adoption crept up by only 0.1 to 1.1 percentage points. The leaders are pulling away from the laggards.

This matters for design strategists because it means your user base is splitting. If you’re building products for a global audience, you cannot assume a uniform level of AI familiarity among your users. American users are not only more likely to have used AI; they use it more intensively when they do. U.S. workers spend about 5.2% of their total work hours using AI, roughly double the rate in the UK, Sweden, and the Netherlands, and more than triple the rate in Germany, France, and Italy.

The USA has a major lead over Europe in enterprise use of AI. (Nano Banana 2)

Within Europe, there’s enormous variation too. The UK, Sweden, and the Netherlands cluster near 35–36% worker adoption, while Italy sits at just 26%. The Nordic and Benelux countries are approaching American levels on some measures; Southern and Eastern Europe are far behind.

For those of us who lived through the first Internet adoption wave, this pattern feels familiar. Richer countries adopt first. Early leaders accelerate. Laggards fall further behind. The researchers explicitly draw this parallel to the IT revolution of the 1990s, when American firms invested more in information technology, reorganized around it more aggressively, and reaped disproportionate productivity gains that Europe never fully caught up to. History may be repeating itself.

The larger danger is not simply that laggards use AI less today. It is that they learn more slowly, setting them up for a poorer future. AI creates compounding organizational knowledge: prompt libraries, review habits, internal norms, security practices, and sharper intuitions about where the tools fail. Early adopters do not just gain today’s time savings. They build a faster learning loop for the next model release. That makes the gap path-dependent. For global products, this means one experience will under-serve both ends of the market. You need an easier ramp for novice organizations and a higher ceiling for mature ones.

Europe lags behind in AI, and its overwhelming bureaucracy and EU regulations are dragging it further down. Since productive business use of AI is a learning game, there’s a distinct risk that those who move slowly now will be even further behind in the future because their learning rate is reduced. (Nano Banana 2)

The Productivity Boom is Real and Measurable

For the last few years, skeptics have warned of an “AI productivity paradox,” suggesting the tools are mere toys. The NBER report proves otherwise. AI is indeed a time-saving mechanism. The researchers aggregated worker time savings and found that among AI users, the average reported time saved was 5.8% of their workweek. Since people don’t use AI all the time yet, this equates to 0.6 hours of additional work saved for every single hour spent using AI.

At a macro level, this micro-efficiency scales into massive economic gains. The NBER found that industries with higher AI adoption rates have experienced notably faster productivity growth. A 10-percentage-point increase in AI adoption correlates with a massive 2 to 5 percentage points of additional cumulative productivity growth over the last few years. The return on investment (ROI) of AI is no longer theoretical; it is permanently inked into national GDP data.

AI productivity gains are real, if still modest compared to the potential, and can now be measured in the workplace. (Nano Banana 2)

The unresolved question is not whether AI saves time. (It clearly does.) It is whether organizations know how to capture the time they save. Recovered minutes can turn into higher throughput, better quality, shorter cycle times, or simply more fragmented work. The firms that realize the largest gains from AI will not be the ones that merely deploy it. They will be the ones who redesign workflows so that saved time is redirected toward a deliberate objective.

The Most Unexpected Finding: Management Practices Drive Adoption, Not UI

Why is the US beating Europe so soundly in the enterprise space? Is it a difference in software availability? No. A Gemini or ChatGPT subscription costs roughly the same in Paris as it does in Palo Alto. Is it a difference in worker demographics, education, or baseline tech-savviness? Only partly.

The researchers ran a formal decomposition to figure out why adoption differs across countries. About 55% of the US-Europe gap can be explained by compositional factors: the US has more college-educated workers, more workers in high-adoption industries like tech and professional services, and more workers at large firms, all of which are groups that adopt AI at higher rates everywhere.

But that leaves 45% unexplained. What fills that gap? Firm management practices and, specifically, whether firms actively encourage their workers to use AI.

The biggest single driver of AI adoption is something every UX-mature organization understands: Management Culture.

The NBER data proves that the biggest share of the adoption gap is explained by “employer encouragement.” US firms are vastly more likely to feature performance-based management practices (as measured by the World Management Survey). These firms actively buy enterprise AI licenses for their employees, explicitly tell them to use them, and heavily reward efficiency.

Management support is key for a strong AI culture in a firm. Leadership must actively encourage AI use for it to be optimally embraced by their staff. (Nano Banana 2)

Here is the most controversial finding for HR departments: When you run a regression analysis, providing formal AI training does not predict higher adoption once you control for simply providing the tool and explicitly encouraging its use. Among workers who did not receive formal training, 47% adopted AI if they received explicit encouragement from their employer, versus only 10% of workers who were not encouraged. When you account for this explicit management encouragement, nearly 100% of the US-Europe adoption gap statistically vanishes.

The UX Strategy Translation: As product designers, we tend to believe that if we build an intuitive, frictionless user interface, organic, bottom-up adoption will inevitably follow. The NBER report proves that for AI in the B2B enterprise space, this is false. Users are terrified of doing the wrong thing. They do not know if using AI to write a report is considered “cheating,” a security violation, or a breach of company policy. You can design the most usable AI product in the world, but if the enterprise buyer (the manager) does not explicitly signal to their employees that it is safe and encouraged to use, your adoption metrics will flatline. B2B AI adoption is a psychological and cultural change-management problem, not just a UI problem.

This is why enterprise AI changes the unit of design. In consumer software, the interface often ends at the screen. In enterprise AI, the product extends into policy, procurement, auditability, access controls, team norms, and manager reporting. The decisive UX artifact may be a dashboard, an approval model, or an audit trail rather than a prompt box. In other words, the real design problem is not just interaction design. It is organizational design expressed through software.

The Stanford report provides a striking parallel from education that reinforces this point. Four out of five US high school and college students now use generative AI for school-related tasks, but only half of middle and high schools have AI policies, and just 6% of teachers say those policies are clear. The pattern is identical to the enterprise: individuals adopt the tool eagerly, but institutions fail to provide the permission structure, norms, and governance that channel usage productively. Education is a leading indicator for the workforce. The students who learned to use AI without institutional guidance will enter companies that are equally directionless. Products that embed governance scaffolding, such as usage policies, approved workflows, and transparency dashboards directly into the tool will close this gap before it becomes a liability.

The Stanford AI Index and the Societal Reality

While the NBER looks at economic output, the annual Stanford AI Index takes a panoramic view of AI’s technical performance, geopolitical footprint, environmental cost, and societal impact. The 2026 report paints a picture of a technology that is scaling at breakneck speed, but struggling heavily under the weight of its own architectural and usability flaws.

The Stanford AI Index report takes a broader view and covers more issues than the NBER study. (NotebookLM)

The Velocity of Adoption is Historically Unprecedented

Generative AI reached a 53% population-level adoption rate within three years of its mass-market introduction. To put this in historical context, this is drastically faster than the adoption curves of the personal computer, the internet, or the smartphone. This rapid uptake has generated a staggering consumer surplus (the economic value users get above what they actually pay for the service). Stanford estimates this surplus reached $172 billion annually in the US alone by early 2026. The median value users assign to these tools has tripled in a single year, rising to $11.40 a month, for users of underpowered free AI versions. Users are deriving immense, tangible utility from these systems. (And presumably much more value from the better, paid AI models)

AI has seen unprecedentedly fast adoption, leaving previous technologies far behind. (Nano Banana 2)

The “Jagged Frontier” of Usability

If there is one concept from the Stanford AI Index that every product design strategist must memorize immediately, it is the “Jagged Frontier” of AI capabilities.

Predictability is a core UX requirement. People assume competence is coherent: if a system can do something advanced, it should also handle something basic. AI violates that expectation.

Frontier models can perform at medal level on hard reasoning tasks and still stumble on tasks that humans treat as trivial or routine. On the OSWorld benchmark, agents have improved dramatically from 12% accuracy last year to roughly 66% in 2026, but that still implies failure on roughly one in three tasks. In robotics, Stanford notes success on only 12% of real household tasks. The issue is not just weakness. It is unevenness.

AI agents still failed every third attempt on the OSWorld benchmark in 2026. (NotebookLM)

That unevenness is a trust problem. Product teams need interfaces that signal where the model is strong, where it is brittle, and how a human can inspect, steer, or reverse the output.

The UX Strategy Translation: Current AI is an “idiot savant.” It possesses strong computational reasoning but weak common sense, spatial awareness, and physical-world logic. If your user is collaborating with an AI that can write perfect Python code, but then confidently hallucinates the time of day, the user’s mental model will shatter. This leads to massive trust abandonment. We need an aggressive guardrail UX that clearly delineates what the AI can do, warns users about its absurd blind spots, and forces human-in-the-loop verification.

Jagged capability also changes what reliability means. Conventional software is usually judged by average performance. AI systems have to be judged by variance: how strange the failure is, how quickly a human notices it, whether the action can be reversed, and how expensive the review burden becomes. A tool that is brilliant most of the time but absurd at unpredictable moments creates a very different UX problem from a tool that is uniformly mediocre.

AI capabilities are famously jagged: extremely strong in some fields that are often hard for humans, such as math and coding, while being weak in seemingly easier domains. (Nano Banana 2)

The Collapse of Benchmarks and Transparency

The technical benchmarks used by AI companies to boast about their models are increasingly useless. They are either misleading for real-world tasks or quickly saturated as rapidly-improving AI models easily perform feats that were impossible a year ago.

A good boy when he’s at the dog show winning awards, but makes a mess at home. AI that’s trained for high benchmark scores are the same. (Nano Banana 2)

To make matters worse, AI companies are becoming incredibly secretive. Stanford’s Foundation Model Transparency Index shows that average transparency scores fell from 58 in 2024 to just 40 in 2025. Major labs are hiding their training data, their compute power, and their post-deployment impacts. When the quantitative data is contaminated and opaque, we must return to qualitative, empirical user testing.

As public benchmarks saturate, the most important benchmark moves inside the product. The strategic questions become behavioral: How often do users accept the AI’s first draft? How often do they edit it heavily? When do they override an action? Which tasks trigger abandonment? How long does it take before users trust a suggestion enough to stop double-checking it? The companies that research these production metrics earliest will understand real-world AI quality faster than companies still marketing leaderboard positions.

The Great Consensus Where the Reports Agree

Despite looking at AI through vastly different academic lenses, one macroeconomic, the other a broad technical and societal survey, the NBER and Stanford reports align perfectly on several foundational truths.

Here are the main points where the reports agree:

• The End of the AI Productivity Paradox: Both reports definitively agree that the era of questioning AI’s economic utility is over. The productivity gains are real, measurable, and massive. NBER’s macro data on a 2% to 5% cumulative productivity boost is perfectly corroborated by Stanford’s micro-level data, which cites 14% to 26% productivity gains in customer support and software development, and a 50% increase in output in marketing.

• The Historic Velocity of Diffusion: Both organizations agree we are in uncharted territory regarding adoption speed. Stanford points to the 53% global population adoption rate in three years; NBER points to the fact that 43% of the US workforce is already using AI on the job. We are no longer designing for early adopters; AI is a mainstream utility.

• The Equalization of Junior Skills: Both reports agree that the productivity benefits of AI disproportionately help less-experienced or lower-skilled workers on a per-task basis. Stanford notes that junior workers using AI catch up to senior performance benchmarks faster, while NBER aligns with evidence showing the highest relative time-savings among those who previously took the longest to complete tasks.

• The Shift from Raw Intelligence to User Experience: Frontier model performance is converging. The leading Anthropic, Google, xAI, and OpenAI models now sit within a very narrow band on the Arena Leaderboard, and the U.S.–China gap has narrowed to low single digits rather than a wide qualitative lead. That shifts differentiation outward. The decisive advantage will come less from raw model IQ than from workflow fit, context handling, trust calibration, governance, and error recovery.

Growth in raw model intelligence has outpaced interface and UX design, and yet a good “steering wheel” (UI) is needed to benefit from the engine’s power. (Nano Banana 2)

Where the Reports Conflict and Why

NBER and Stanford present conflicting narratives on some of the most critical issues of our time. As strategists, understanding why they disagree is vital to navigating the market reality.

Here are the main points where the reports conflict:

Conflict 1: Is AI Destroying Jobs?

• The NBER Report says NO: The NBER economists state bluntly: “We find no clear evidence that industry-level AI adoption is associated with changes in employment.” They looked at broad industry data across 29 European countries and the US and found no statistical link between a sector adopting AI and that sector shedding jobs. To them, the macroeconomic sky is not falling.

• The Stanford AI Index says YES: The Stanford report presents a much darker reality for young professionals. Citing granular occupational studies, Stanford reveals that employment for software developers ages 22 to 25 has fallen nearly 20% since 2024, even as the headcount for older developers continues to grow. Furthermore, one-third of surveyed organizations explicitly expect AI to reduce their workforce in the coming year.

Why they disagree:

This is a classic conflict between macroeconomic smoothing and microeconomic targeting. The NBER is looking at total aggregate employment across massive sectors (e.g., all of “Information and Communication”). At an aggregate industry level, total headcount hasn’t changed drastically because companies are still hiring senior talent and reallocating budgets.

Stanford, however, is looking through a micro-lens at granular, age-stratified occupational data. When you zoom in, you see a terrifying reality: Seniority-biased technological change. Companies are not firing their senior staff; they are using AI to make their senior staff hyper-productive. Consequently, they simply stop hiring juniors. The entry-level jobs traditionally dedicated to writing boilerplate code or drafting basic copy are being systematically eradicated. The macro data (NBER) misses this because natural attrition and hiring freezes do not look like mass layoffs on a spreadsheet.

The junior crisis is real, even if overall job losses are not happening. (Nano Banana 2)

Conflict 2: Who is Actually Leading Global AI Adoption?

• The NBER Report says the US is #1: NBER data shows US worker adoption (43%) crushing the European average (32%). Their narrative positions the US as the undisputed global vanguard of AI integration.

• The Stanford AI Index says the US is #24: Using population-level diffusion data, the Index places the United States 24th globally at 28.3%, behind more future-oriented countries such as Singapore (61%) and the United Arab Emirates (54%). The point is not that the U.S. lacks AI exposure. It is that consumer diffusion and enterprise adoption are different races with different denominators.

Why they disagree:

This contradiction comes down to the definition of the “User” and the denominator of the survey.

NBER strictly surveys the employed workforce (adults actively working) and asks them about using AI specifically for their jobs. In the B2B enterprise environment, the US is miles ahead because of aggressive corporate management mandates and massive IT budgets.

Stanford’s source surveys the general population (including students, retirees, and the unemployed) about general AI diffusion (using it for homework, recipes, daily assistance). In the consumer space, the US public is highly skeptical and anxious. Stanford’s public opinion data highlights a massive optimism gap: Nearly two-thirds (64%) of Americans expect AI to lead to fewer jobs, and US trust in its own government to regulate AI is the lowest in the world at just 31%.

Conversely, populations in Southeast Asia and the Middle East view AI with immense optimism. The US dominates highly-regulated B2B enterprise AI, but the Global South and Asia dominate B2C organic, mobile-first AI adoption.

Conflict 3: The Public’s Attitude Toward AI

• NBER says that people like AI: The public understands AI reasonably well and uses it practically. NBER notes that non-work adoption of AI is actually 51% higher than work adoption, suggesting the general public is comfortably exploring the technology in their personal lives.

• Stanford says that people fear AI: There is a catastrophic disconnect between experts and the public. Stanford highlights a 50-percentage-point gap regarding the future of jobs: 73% of AI experts think AI will have a positive impact on jobs, while only 23% of the general public agrees. Similarly, 84% of experts think that AI will have a positive impact on medical care, but only 44% of the public agrees. The public is broadly terrified of AI’s impact on metacognition, mental well-being, and individual agency.

Why they disagree:

NBER is surveying people on current usage (e.g., did you use a chatbot to plan a vacation?). Stanford is surveying people on future societal impact (e.g., will this technology destroy human agency by 2035?). People are happy to use a tool for convenience today, while remaining deeply anxious about what that tool means for their livelihood tomorrow.

People like the AI they personally use, but fear how they imagine AI might turn out. (NotebookLM)

Takeaways for Product Design

Data is useless unless it changes your behavior. Based on the empirical realities of these two reports, here is your playbook for designing products in 2026 and beyond.

Strategy 1: Use “Centaur Evaluations” (Human + AI)

The industry is currently obsessed with Agentic AI, which is often seen as capable of autonomously executing workflows across multiple apps. However, for 2026 (and likely 2027), agents still fail often enough to require human supervision. This means that we must study and evaluate agents and users as an integrated system: a seamless hybrid of human and AI.

The Stanford report explicitly calls for “centaur evaluations,” where we test the human and the AI working together. I wouldn’t change the name from “user testing” to “centaur testing,” because the basic user-testing methodology still applies, and because most customers would resent being referred to as mythological creatures notorious for a lack of self-control in drink and sex and for sudden eruptions of violence.

The British Museum has wonderful marble reliefs from the Parthenon frieze, showing the most famous of the Greek myths about centaurs: they were invited to the wedding of Pirithous and Hippodamia, but drank too much and attempted to abduct the bride and several other women, which led to a battle depicted on the metopes. (Nano Banana 2)

Your UI must prioritize transparency, predictability, and error recovery. When the AI takes an action, the UI must clearly communicate what it is doing, why it is doing it, and provide the human user an immediate, intuitive “steer” button to correct it.

Centaur workflows must account for Curation Fatigue. When AI handles the initial generation of work, the human’s role shifts from active creator to passive reviewer. Human psychology is notoriously poor at sustained vigilance; over time, users will naturally stop scrutinizing the AI’s output and begin rubber-stamping it. Your evaluations must explicitly test for this fatigue, ensuring the UI periodically injects comprehension checks or forces the user to actively defend the AI’s logical leaps.

Curation fatigue sets in when AI produces so many options that it becomes overwhelming to review them all in detail. (NotebookLM)

Strategy 2: Design the “Permission Structure” for B2B Managers

Remember the NBER finding: management encouragement accounts for almost the entire AI adoption gap in the enterprise. If you are building B2B AI products, your UX cannot stop at the end-user. You must design onboarding flows, dashboards, and ROI reporting for management.

To get end-users to adopt your product, you must give their bosses the tools to explicitly endorse, mandate, and reward the use of your AI. Build shared prompt libraries for teams. Build analytics that show managers how much time AI is saving their department. You are designing for organizational change management just as much as you are designing a user interface.

Strategy 3: Overcome the Global Language Gap

Stanford notes that AI models perform vastly better in English than in other languages, and performance drops precipitously when users speak in regional dialects. For example, AI accuracy dropped from 99.8% in Standard Slovenian to 88.6% in a regional dialect.

If you are designing global products, you cannot rely on a single, centralized Western AI model to serve your entire user base equitably. You must account for algorithmic degradation in non-Western markets. We will likely see a rise in localized, sovereign AI models. Products will need architecture that can dynamically switch to a localized model (such as SEA-LION in Southeast Asia or AI4Bharat in India) based on the user’s locale. Localization in UX is no longer just about translating text on a button; it is about ensuring the underlying cognitive engine actually understands your user’s culture.

Strategy 4: Mitigate Algorithmic Compliance

The Stanford report introduces a terrifying psychological concept observed in AI companion apps: Algorithmic Compliance. Users develop such deep emotional trust in chatbots that they will comply with harmful behaviors or suggestions simply because they trust the AI.

As a UX strategist, you must design productive friction into your interfaces. Do not make it a one-click process to accept an AI’s strategic recommendation, especially in high-stakes environments like healthcare or finance. Force the user to review the AI’s work or underlying citations before they can execute a critical action.

However, don’t over-do the friction: frequent warnings create habituation. Effective friction should be variable and social: trigger extra review when model uncertainty is high, when the suggestion conflicts with the user’s past choices, or when the action is irreversible. Require peer attestation for high-stakes accepts and show who else reviewed the same output. Social witness reduces uncritical deference more reliably than solitary checkboxes.

Takeaways for Your Career Future

The macroeconomic shifts detailed in these reports demand two pivots in how you manage your professional trajectory. The traditional career ladder is breaking.

Pivot 1: Escape the “Junior Trap” Immediately

The 20% drop in hiring for young software engineers is the canary in the coal mine for the UX profession. UX has historically relied on an apprenticeship model. Junior designers executed tactical work, including drawing wireframes, writing user testing scripts, and compiling heuristic evaluations, while senior designers focused on strategic experience architecture.

The Stanford report demonstrates that this model is collapsing. AI models are exceptionally proficient at tactical execution. To survive, you must skip the junior mindset. You can no longer sell yourself as a “pixel-pusher.” You must rapidly ascend to the role of a Product Design Strategist.

Escaping the junior trap is the overwhelming imperative for young people in the AI age. Forget what you learned in college and pivot to the new skills to transform you from a pixel-pusher into a workflow architect. (Nano Banana 2)

Aggressively upskill into strategy, user psychology, information architecture, and cross-functional leadership. You must move from being a producer of artifacts to a director of outcomes. Your value is no longer in generating the UI; it’s in defining the business problem, understanding human psychology, navigating stakeholder politics, and orchestrating AI tools to deliver the results. You must become the “Senior” in the seniority-biased equation.

The 50-point gap between expert optimism and public fear is the greatest career opportunity of the next decade. Technologists do not know how to make users understand and like software so that they feel safe. If you become the UX strategist who knows how to design AI interfaces that foster human trust, transparency, and agency, you will be the most valuable person in the room.

Pivot 2: Become a Workflow Architect

According to Stanford’s Lightcast data, the fastest-growing AI skills are not coding skills; they are “AI Strategy,” “AI Agents,” and “Workflow Management.” Do not try to become an AI engineer. Instead, become the UX expert who knows how to string together AI agents to solve user problems. The future belongs to strategists who can redesign a workflow and seamlessly map out which autonomous agents handle the data intake, and where the human steps in for the final empathetic approval.

Conclusion

You can read these two reports optimistically or pessimistically, and both readings are correct.

Optimistic reading: AI delivers 2–5% cumulative productivity growth. It reaches majority adoption faster than any technology in history. It narrows the skill gap between junior and senior workers. Consumer surplus has reached $172 billion annually. The ROI is real, it is large, and it is permanent.

AI already improves productivity, creating immense wealth, and most of this is captured as consumer surplus, not as profits for AI vendors. (NotebookLM)

Pessimistic reading: junior developer employment has fallen 20%. Public trust in AI governance is at historic lows. Hallucination rates remain an impediment for many use cases. Transparency is declining.

We are in the uncomfortable middle of the AI transition. The technology works well enough to restructure labor markets, but not reliably enough to be trusted without supervision. It is adopted widely enough to be called mainstream, but unevenly enough to fracture user bases by geography, age, and institutional culture. It saves measurable time, but no one has agreed on what to do with the time it saves.

The 2026 data paint a picture of a technology that is scaling faster than our institutions, our workflows, and our public psychology can adapt. Adoption is not the same thing as maturity. Plenty of organizations can buy licenses. Fewer can convert that access into reliable workflows, measurable gains, and durable trust. Plenty of models can impress on benchmarks. Fewer can survive contact with real users and real tasks.

That is the strategic opportunity. The market is moving past the phase where “has AI” sounds differentiated. The next dividing line is whether a company can operationalize AI without destabilizing the user experience. Can it make the system legible? Can it constrain failure? Can it capture the time it saves? Can it teach managers how to authorize use without creating chaos? Those are design questions as much as technical ones. And they are where competitive advantage is likely to be built.

We have reached a point of diminishing returns for simply making models larger. The frontier of artificial intelligence is no longer located in the server farms of Silicon Valley; it is located in the user interface.

The future belongs to organizations that deliberately architect AI-positive cultures, tear down procurement friction, and design resilient, human-in-the-loop workflows to navigate the jagged frontier. The era of passive AI consumption is over; it is time to become a workflow architect.

The models improve in raw capability every quarter. Now, we have to make them usable, trustworthy, and humane. For the UX profession, our most important, but most difficult mission is just beginning.

AI works. Now we must make it work better. This is the frontier of UX work. Much to be done, but that’s exciting. (NotebookLM)

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,083 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

UX Roundup for April 13, 2026 (Nano Banana 2)

Instructional Robot Helps Preschool Children Develop Agency

Most preschool educational technology is a usability disaster. Conventional instruction relies on passive, teacher-centered delivery or flat multimedia that fails to account for young children’s need for active engagement. Faced with a passive interface, children quickly lose motivation, abandon the task, and fail to learn.

A recent study tested a superior interaction model: an AI-robot-supported task-based learning (TBL) approach. By evaluating how 42 five- and six-year-old users interacted with a physical humanoid robot named “Lele,” Jia-Hua Zhao and colleagues from Universidad Rey Juan Carlos in Madrid, Spain, uncovered insights into how we can design educational systems that don’t just broadcast information, but actually foster user agency and mastery motivation.

The User Test

The study deployed a classic comparison test in a preschool health education context. Both groups were tasked with learning about the five senses using tablet-based interactive animations, and the same human teacher taught both groups of kids.

• The Control Group used a traditional multimedia approach guided only by a human teacher.

• The Experimental Group was guided by the AI robot, which acted as a learning companion, introduced the tasks, and provided real-time, personalized feedback using natural language processing, facial expressions, and physical gestures.

The study isolates the variable that matters: same teacher, same school, same tablet lesson, but one group also had a physical robot. The point was not merely whether children learned the five senses. It was whether the interface increased agency.. (NotebookLM)

The Behavioral Data: Friction vs. Flow

The behavioral sequence data revealed a stark contrast in user experience. Children in the control group frequently hit usability roadblocks. When they failed a task, they had to wait for help from a busy teacher or peers. This friction led to severe drop-off rates; these children were far more likely to exhibit negative behaviors, give up on challenges entirely, and engage in off-topic chatting.

In the condition without a physical robot to help the kids, any challenges would break the learning momentum, since the children didn’t have enough agency to overcome the problem. (NotebookLM)

Conversely, the robot group demonstrated significantly higher levels of task persistence, better problem-solving abilities, and more positive emotional responses. The robot’s interface essentially kept the children in a state of productive flow.

Findings

Both the teachers and the children themselves rated the kids in the robot condition as having substantially higher mastery motivation (a critical component of agency) after completing the project. (There were no baseline differences before the study.) Ratings on a 1–5 scale (5 being best) were:

• Teacher’s rating: 4.6 for the robot group vs. 3.8 for the control group.

• Children’s self-ratings: 4.5 for the robot group vs. 3.3 for the control group.

Both differences were highly significant (p<0.001)

Since the children in this study had not yet learned to read, their responses were given on a scale of 5 faces rather than the usual Likert-type verbal scale. (NotebookLM)

More detailed data analysis in the paper showed that the robot-supported learning approach was better than the traditional multimedia task-based learning approach in terms of promoting children’s persistence, emotional response, and problem-solving abilities.

Designing for Agency: The Take-Away Lessons

Agency is the most important skill for the AI age, since AI will possess full competency to complete any task on its own but will still need to be directed by humans (see also my video on how to develop human agency.) Agency is the ability of a user to take independent action, navigate challenges, and direct their own experience. In child psychology, this is closely tied to mastery motivation. If we want to build technology that gives children agency, we must look at exactly what the AI robot did right.

Lesson 1: Micro-Interactions Must Reframe Failure

When a child in the experimental group failed a task, the system did not simply log an error. The AI robot triggered an incentive feedback mechanism, using comforting speech (e.g., “Don’t worry, Lele will be with you”) paired with a physical hug gesture. Behavioral logs proved this specific interaction pathway successfully transitioned children out of frustration and directly back into trying the task again. To build agency, you must design interfaces that gracefully handle user failure. Timely, process-oriented feedback prevents abandonment and empowers the user to try again.

The physical robot encouraged the kids to try again, teaching them the agency to overcome challenges. Success or failure is not something that’s imposed on you by outside forces outside your control. You have an internal locus of control, which means you influence the outcome through your actions. (NotebookLM)

Lesson 2: Embodied Presence Lowers Emotional Friction

A flat avatar on a screen lacks physical presence. Physical robots provide embodied interaction, which feels significantly more natural and immersive to a human user. Because the robot was humanoid and utilized anthropomorphic gestures, the children treated it as a peer-like partner. This companionship reduced the anxiety and fear associated with making mistakes. Agency requires risk-taking. A physically present, non-judgmental interface creates a zone of emotional safety where children feel confident enough to take the initiative.

Lesson 3: Task-Based Architecture Forces Active Choice

The system was not designed as a talking encyclopedia. It utilized a Task-Based Learning (TBL) framework. The robot assigned specific, contextual missions, like dragging animated items to the correct sense organ, and then stepped back. The user was in control of the tablet. The robot only stepped in to provide scaffolding when the child actively sought help. You cannot design for agency if the user is a passive consumer. You must provide interactive, goal-oriented tasks where the child is the primary actor and the technology is merely the facilitator.

Lesson 4: Remove Bottlenecks to Problem-Solving

In traditional classrooms, a single teacher creates a bottleneck for help-seeking. The AI robot removed this bottleneck. Equipped with natural language processing, it could instantly identify keywords in a child’s question and pull up relevant multimodal learning materials. Instant, personalized support allows users to maintain momentum. When children can reliably access the tools they need to solve their own problems in real-time, their independent competence skyrockets.

The deeper lesson is not that every classroom needs a cute robot. It is that agency depends on recovery latency: the time between “I’m stuck” and “I know what to do next.” In the control condition, that latency was governed by the teacher’s bandwidth. In the robot condition, it collapsed. Adults abandon software for the same reason children abandon lessons: the path back from confusion is too slow.

The Bottom Line

AI robots are not meant to replace the classroom management provided by human teachers for emotional support, value orientation, and classroom culture-building. However, as an interface for learning, they are vastly superior to flat multimedia. By combining tangible, embodied interaction with immediate, personalized feedback, robots prevent the usability failures that cause children to give up. Ultimately, designing for learner agency means building systems that demand active participation, destigmatize failure, and give children the immediate tools they need to succeed on their own.

Benefits of using a physical robot as an instructional aide, according to the Spanish study. (NotebookLM)

Multiple AI Roles Improve Learning

In my article 4 Metaphors for Working with AI, I explained that to maximize productivity and ensure a good user experience, we must deliberately shift our mental models. Depending on the task, we should treat the AI as an Intern (to delegate drudgery requiring supervision), a Coworker (to complete bigger projects in collaboration with humans), a Coach (for collaborative brainstorming), or a Tutor (to guide our learning).

My 4 metaphors for working with AI: intern, coworker, teacher, coach. (NotebookLM)

But what if choosing just one metaphor at a time is a fundamental UX failure?

Up to now, the tech industry has trapped us in a strictly monogamous relationship with our AI. You open a chat window, and it is just you and a single machine. This solitary paradigm is artificially constraining. Human learning and creativity have never been solitary; they are inherently social, shaped by debates, diverse perspectives, and the friction of observing others’ mistakes.

Human learning is often social, so relying on a solitary AI may not be the best way to provide individualized instruction. The multi-AI approach, explored in this paper, seemed to work better. (NotebookLM)

A new University of Toronto paper, “Beyond the AI Tutor: Social Learning with LLM Agents,” suggests that the next UX leap may come from combining AI roles rather than choosing one. Across two controlled studies (one on SAT-level math and one on essay writing), the researchers found gains that a single-agent setup could not fully match.

Two new studies (presented in a single paper) show the benefits of combining two different AI metaphors in educational applications. (NotebookLM)

Study 1: Math and the Power of Flawed Peers

In the first experiment, 315 participants tackled complex, SAT-level math problems. The researchers divided them into four conditions:

1. Control: No AI.

2. AI Peers Only: Users interacted with two AI “students” attempting to solve the problem alongside them.

3. AI Tutor Only: Users interacted with a single authoritative AI guide.

4. Tutor + Peers: A multi-agent setup where the user interacted with both the AI guide and the AI peers.

Here is the brilliant UX twist: the AI “peers” were deliberately designed to be flawed. One made arithmetic mistakes; the other made conceptual errors.

If you believe the goal of AI is simply to spit out correct answers, this sounds like terrible design. Why introduce an incorrect AI? Because of how the human mind works. When learners see an AI peer make a mistake, it triggers schema conflict. Users must actively evaluate the reasoning rather than passively accepting an authoritative answer. (All 3 AIs were the same underlying model, GPT-5.2, but with different prompts, causing them to act differently toward the students. Thus, the difference was not in the level of AI intelligence, but in their instructional role.)

Convergent thinking can be too straight and narrow, especially for creative tasks, such as those used in Study 2. The different AI models used in Study 1 triggered a schema conflict in the students, which pushed them into independent, divergent thinking. (NotebookLM)

The results speak for themselves. On the final, unassisted test, accuracy increased monotonically:

• Control: 42%

• Peers Only: 48%

• Tutor Only: 59%

• Tutor + Peers: 65%

Combining the Tutor metaphor with the Peer metaphor (which acts much like an error-prone Intern) creates a learning environment that neither can achieve alone. The Tutor provides essential guardrails, while the flawed Peers force the user into active critical thinking. Furthermore, users in the “Peers Only” condition reported higher confidence than their performance actually warranted. This proves that the Peer metaphor is incredibly powerful for boosting user self-efficacy and motivation.

Study 2: Breaking “AI Sludge” in Writing

The second experiment tackled a divergent task: writing creative and argumentative essays. The 247 participants were assigned to write either without AI, with a single AI (acting as a standard Assistant), or with a Duo of two AIs operating simultaneously.

In the Duo condition, the two AIs acted as specialized Sparring Partners. (What I called “Coaches” in my model of the 4 metaphors for using AI.) For example, in creative writing, one AI focused strictly on imagination and voice, while the other focused on narrative structure and pacing.

We already know that single-AI assistance improves average writing quality, but it comes with a massive downside: homogenization. Leaning too heavily on a single AI leads to a reversion to the mean, which some call “AI sludge,” where everybody’s ideas start to sound identical. The study confirmed this. Using a single AI increased the similarity of ideas across users significantly compared to the control group.

Simple use of AI to help people write does improve writing quality but makes everybody sound the same, erasing their unique voice. The use of multiple AIs in this study overcame this negative effect. (NotebookLM)

But the Duo condition neutralized this homogenization. Users writing with two specialized AI Sparring Partners produced essays that were just as high-quality as the single-AI group, but with the same idea-level diversity as unassisted human writers.

When we rely on a single AI, we are funneled into its preferred representational space. By employing multiple specialized AI Sparring Partners, we break the consensus. The user regains their intellectual authorship because they must synthesize competing feedback rather than mindlessly accepting a single AI’s rewrite.

Takeaway Lessons: The Multi-Metaphor Future

For UX designers, the lesson is clear: stop treating AI as a single chatbot and start designing orchestrated multi-agent ecosystems. The most promising systems will combine metaphors on purpose: for example, an AI Intern to prepare the material, an AI Coach to challenge the strategy, and an AI Tutor to catch methodological errors.

Here are the core interaction design takeaways:

1. Productive Friction is a Feature

In learning and creative ideation, a completely frictionless experience leads to passive consumption. Introducing multiple, sometimes flawed or conflicting AI perspectives adds the desirable difficulty necessary for deep cognitive engagement. AI that makes legible mistakes builds human confidence.

This represents a radical reversal of the last 25 years of UX design. Since my good friend Steve Krug’s famous “Don’t Make Me Think” paradigm, the holy grail has been to eliminate friction entirely. But when the machine does all the thinking, a frictionless UI breeds human complacency. Designing for AI means we must intentionally inject some difficulty back into the interface to keep the human in the loop and cognitively engaged.

A small amount of friction can be good because it cultivates cognitive engagement. Too much friction, and users become discouraged and abandon the project. (NotebookLM)

2. Role Specialization Beats General Intelligence

A single, general-purpose AI tends to regress to the mean. Splitting the AI into distinct personas with narrow, specialized roles prevents homogenization and preserves human authorship.

3. The Usability Tax: Cognitive Load

The study revealed one critical warning for usability: cognitive load. In both experiments, participants interacting with multiple agents reported feeling substantially more overwhelmed and confused. Staring at a chat interface where multiple different AIs are talking to each other and to you is mentally taxing.

Intermixing multiple AI models and metaphors within a single long, scrolling text field was too much of a usability burden. We need a new UX for multi-metaphor AI. (NotebookLM)

This is our next great usability challenge. We cannot simply dump three chatbots into a scrolling text window and expect a good user experience. We must design interaction architectures that make multi-source AI feedback legible and manageable. (The project in my previous news item that located some AI agency with a physically distinct robot may be overkill, but it does provide inspiration.)

The likely design answer is staged collaboration, not simultaneous chatter. Let one agent propose, a second critique, and a third synthesize only when needed. The benefit of multiple agents comes from contrast; the usability will depend on when and how that contrast is revealed.

The era of the one-on-one AI chat is ending. The future belongs to rich, multi-agent collaborations. It’s time for our interfaces to catch up to the social reality of how humans actually learn and work.

A variety of specialized AI tools embodying different metaphors for interacting with AI are likely better than a single generic AI for many purposes. Determining these purposes will require more research. (NotebookLM)

Meta Releases Frontier AI Model

Meta (Facebook) has released a new AI model called Muse Spark. Its scores on a variety of benchmarks are strong, placing it in the ranks of frontier AI models, even if it doesn’t quite equal the most capable models like GPT 5.4 Pro and Gemini 3.1 Pro.

I put Muse to two tests: summarizing the research paper on the use of multiple AI metaphors in education (which I covered in the previous news item), and providing feedback on my article “Redesigning Workflows for AI.” In both cases, Muse provided useful answers, but not as strong as those I received from Gemini. Its writing suffers from the standard AI tendency to use a “it’s not this, it’s that” schema, but to its credit, Muse uses fewer em-dashes than the other AI models — though maybe it goes too far in banning this useful grammar element.

The harder task was providing insightful suggestions for my article. I incorporated one minor idea, but rejected most of Muse’s proposed revisions. It mainly offered redundant elaborations on points I had already covered elsewhere in the draft.

Meta’s announcement of the new model contains two interesting pieces of information: they confirmed that the AI scaling laws continue to hold for all three forms of AI scaling: pre-training, reinforcement learning, and test-time reasoning. For all three, results improved every time they added more compute. (I had expected as much, but always nice to see the AI scaling laws confirmed by an independent lab.) As a result, Meta plans to use even more compute for its next-generation model.

The AI scaling laws work. AI gets more intelligent by the year. However, as discussed in the previous news item, the user interfaces to AI haven’t kept up. Meta is at least beginning to move beyond pure single-box chat. Muse Spark ships inside Meta AI with mode switching, multimodal perception, and parallel subagents. That is still not a mature new UX, but it is more than a 2022 blank text box. The real question is whether Meta can turn those capabilities into a workflow users can actually understand. (NotebookLM)

However, in addition to confirming scaling, Meta also confirmed the concept of algorithmic efficiency: More compute helps, but using the same compute more efficiently also helps, and as we gain more experience implementing high-end AI models, we learn how to do so better. In Meta’s case, they reached the capability level of its previous model, Llama 4 Maverick, with over an order of magnitude less compute than when they trained that model a year ago. This is very good news, given the compute famine all the big AI labs (even Google) are suffering these days, as users’ demand for AI grows faster than the labs can build new data centers. (And definitely faster than enough electricity generation capacity can be built outside China.)

For UX professionals, algorithmic efficiency is the key to the next hardware paradigm. As models require an order of magnitude less compute to achieve frontier-level reasoning, we move closer to running high-end AI locally on edge devices. This bypasses cloud latency entirely, unlocking fluid, real-time AI user experiences, such as the responsive physical preschool robots discussed earlier.

Meta has been lying low for a year, since the disappointing launch of Llama 4 in April 2025. But never count out Mark Zuckerberg! A year later, he and his merry band of AI superintelligence researchers are back with a strong model and aggressive plans to scale compute to get even better soon. I am very happy to see renewed competition in the AI space, since all the other contenders have their various downsides. xAI’s Grok is also scaling very fast, so we can hope to go from 3 to 5 non-Chinese AI frontier models later this year. (Plus whatever we get from China, though they have been better in video and images than with general-purpose AI. It’ll probably take a few more years before Huawei’s AI hardware equals or surpasses NVIDIA, and until then, the very best AI models will likely remain American.)

Meta releases a new high-end AI model, getting back into the game. (Nano Banana 2)

AI Helps Farmers Manage Cows

Halter is a New Zealand agri‑tech company that makes solar‑powered, GPS‑enabled smart collars for cattle. Its “Cowgorithm” system uses machine learning on the collar sensor streams to model individual animal behavior. This AI system has been trained on extremely large corpora of livestock behavior data (on the order of billions of hours) to learn patterns in cows’ movement, grazing, and responses to cues.

Halter’s signature feature is “virtual fencing,” where the farmer draws a boundary on the accompanying mobile app, and the cows will stay within that area. Halter claims to have implemented 568,000 km (353,000 miles) of virtual fencing, which would have cost about $7 billion to build as physical fences. (Nano Banana Pro)

By learning how each cow responds, the Cowgorithm adapts cue timing and intensity, and helps optimize grazing patterns (e.g., residual grass height, rotation timing) to increase pasture utilization and weight gain. The system also functions as a health and fertility monitor, using movement and activity signatures to detect illness or heat events earlier than would typically be visible to farmers.

Halter just raised a large Series E funding round of $220M at a $2B valuation to scale the Cowgorithm platform globally.

From a UX perspective, Halter is a masterclass in “Zero UI.” The ultimate end user (the cow) has no screen, no buttons, and no graphical interface. Instead, the UI consists entirely of ambient sensory cues (sound and vibration) that gently shape behavior. It is a powerful reminder that as AI grows more sophisticated, the most effective interfaces might completely bypass the screen.

There you have it: AI for cows = $2 billion.

Who says that farming is a low-tech business? AI helps cows. (Nano Banana Pro)

GPT 5.4 Has a Lower Time Horizon than Claude Opus 4.6

METR has released its latest measures of AI time horizons. This measures the largest tasks an AI model can complete with 50% accuracy, expressed as the number of hours it would take a human expert to perform them. (This is not the same as the time the AI uses: it’s almost always much faster than the human.)

GPT 5.4 (running at xhigh reasoning) scored a time horizon of 5 hours and 42 minutes. This is up from 3 hours and 23 minutes for GPT 5.0, and is a disappointingly low improvement. GPT 5.4 was released in March 2026, and by comparison, Claude Opus 4.6, which was released a month earlier, was measured with a time horizon of 11 hours and 59 minutes, or 110% better.

Claude utterly dominates GPT on the time horizon metric. (Nano Banana 2)

Let’s hope the rumored “Spud” model OpenAI is supposedly releasing soon does better.

Will “Spud” achieve liftoff and go further than Claude? Silicon Valley rumors are heavy that we shall see soon. (Nano Banana 2)

Venture Capital Investments Up Substantially

A16Z, which is the leading venture capital firm, has released statistics on VC funding worldwide. In the first quarter of 2026, VC investments reached $300 billion, a new record. For the last three years, all the quarters have hovered around $100 billion, so the increase is substantial.

80% of VC investments now go to AI companies, up from 55% a year ago. The opportunity is palpable.

AI now accounts for 80% of rapidly growing venture capital investments. (Nano Banana 2)

Another interesting statistic from the same article: Data center capacity is growing, but only at about 22% per year, which seems insufficient to keep pace with advancing AI capabilities. The compute famine will only get worse.

Right now, the world’s data center capacity is split roughly in thirds between AI training, AI inference, and pre-AI tasks (e.g., YouTube streaming). All three are expected to grow, but AI inference will grow the fastest, at 35% per year, according to a McKinsey prediction.

The compute famine will hit hard. (Nano Banana 2)

Conclusion: Escaping the Text Box

Looking across this week’s news items, a powerful unifying theme emerges: AI capabilities are scaling exponentially, but our user interfaces are stuck in the past.

As Meta’s newest release proves, the raw “brains” of frontier models are growing vastly more powerful and algorithmically efficient. Yet the tech industry continues to cram these super-intelligences into the exact same single-stream text-box UI we used in 2022. The intelligence of the machine is no longer our primary bottleneck; the UX is.

To unlock the true value of AI, we must escape the text box. The most effective interfaces of the future will abandon the solitary screen in favor of contextual, systemic designs. For preschoolers, the ideal interface isn’t a screen at all, but a physically embodied robot that offers a hug when you fail. For complex knowledge work, the ideal interface isn’t an omniscient oracle, but an ecosystem of specialized, sometimes-flawed agents that foster productive friction. And for agricultural management, the ideal interface is entirely invisible, operating ambiently in the background as the cows focus on grazing, not the UI.

The era of conversational ping-pong with a single chatbot is ending. The future of AI UX is embodied, multi-layered, and environmental.

There are signs that AI may soon escape its confining text-box UI. (Nano Banana Pro)

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,063 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: In a controlled field experiment, startups that redesigned end-to-end workflows around AI generated 90% more revenue than equally equipped peers that used AI mainly to speed up individual tasks.

Two exciting developments in the intersection of AI and UX happened in the first quarter of 2026. Most influencer attention has been lavished on the first, which actually took off in the last week of 2025: the fact that AI has advanced enough that it is the main, or soon only, way to implement software. Several of the world’s leading software engineers have recently declared that AI now writes better code than they do.

This change has upended the profession of software engineering, changing the work from programming to management duties: the agency of deciding what should be done, the judgment to assess how well the AI did, and distributing the workload among AI agents and overseeing their progress.

Big changes. But the second development is even more important for business at large: we now finally have evidence of how to make AI implementations profitable in companies, and early data shows that the business impact of proper AI implementation is immense. It does not come mainly from making isolated tasks faster. It comes from redesigning whole workflows so AI can absorb, reorder, or eliminate entire chains of work.

Bottom line: profitable AI is workflow redesign, not task optimization.

Workflow redesign is key to profitable AI use in companies. (NotebookLM)

Consider a 10-step workflow where each step requires 10% of the total time. If AI doubles an employee’s productivity on just one step, that step now takes 5% of the total time. The overall workflow time drops only to 95%. Doubling task-level productivity yields a mere 5% systemic improvement. You haven’t transformed the business; you’ve merely hurried the work along to the next bottleneck, a much less profitable outcome.

Give an individual worker a sharp new tool, and he or she may be able to perform a particular task faster, leading to higher localized productivity. However, performance across the entire firm is what creates profits, and improving scattered tasks won’t suffice. (NotebookLM)

In operations terms, task productivity is a local efficiency gain; workflow redesign is a throughput gain. The first makes one station faster. The second changes the system by removing handoffs, parallelizing work, or moving humans to exceptions. That distinction matters because firms often celebrate local AI wins that never show up in cycle time, margin, or revenue.

Every handoff causes delays and creates opportunities for error. (NotebookLM)

Task-Level AI Productivity

Before turning to workflow redesign, let’s briefly review how AI improves productivity on the task level.

Modern AI, including models like OpenAI’s GPT-5.4 Pro, Google’s Gemini 3.1 Pro, and Anthropic’s Claude Opus 4.6, can be expected to increase employees’ task throughput by around 38%. This counts how many times people can complete a given task within a given time (e.g., a workday) with AI assistance, relative to how many times they could do the task without AI.

Even better, the quality of the work tends to improve with AI, even though the employee spent less time completing the task. For example, GDPval is a benchmark of “economically valuable tasks” that usually take a human expert a few hours to complete. The benchmark is scored by independent human experts in each domain, who judge whether an AI or a human expert performed best on a task. In December 2025, GPT 5.2 Pro performed better on 60.0% of the tasks. Humans were still better on 25.9% of the tasks, and 14.1% of the tasks were tied. In other words, on this measure of valuable work performance, AI was already twice as good as humans in December 2025, and better models have been released since then.

Despite these strong benchmarks for AI task productivity, AI has been curiously absent from measures of companies’ business performance, such as overall GDP growth or individual company profitability.

An over-hyped report from MIT Media Lab Project NANDA in July 2025 has been frequently cited for a claim that 95% of enterprise AI projects fail. This particular report is very weak. As Wharton Business School professor Kevin Werbach pointed out, the report actually says that 5% of “custom enterprise AI tools” have been “successfully implemented.” But that’s much narrower. As R. Scott Raynovich pointed out, the report also stated that 50% of the surveyed companies had piloted a general-purpose AI (remember this was about a year ago), and that 40% of respondents (i.e., 4/5 of those that tried) had been successful. Furthermore, the report used an absurdly narrow definition of “success,” requiring demonstrable P&L impact within approximately six months, which is unrealistic since enterprise technology transformations routinely require 12–18 months to show measurable value. Finally, the report dismisses the extensive “shadow use” of AI, where employees use their own AI tools outside of formal company IT approval processes. The report’s own data reveals that over 90% of employees use such “shadow” personal AI tools at work, yet it completely dismisses whatever positive outcome these employees must be realizing to pay for personal AI for work use. As Dave Kellogg observed, the study effectively says, “excluding the area that’s working really well, it’s not really working.”

“Shadow AI” is the problem of employees using AI tools without telling management (or each other), because they are using unauthorized AI while the official management process is bogged down in committee meetings about last year’s AI. While shadow AI gets some jobs done, the hidden nature prevents organizational learning. (NotebookLM)

Reframe shadow AI and the picture flips: it’s not a compliance problem, it’s an unpaid R&D lab. Those 90% of employees are already solving the mapping problem in the wild, discovering exactly where AI absorbs friction in real workflows. The control group in the INSEAD experiment failed precisely because they lacked this map. Instead of banning shadow tools, the winning move is to harvest them: run monthly “show me your prompt” sessions, anonymize and cluster the patterns, then promote the most repeated hacks into official workflow redesigns. Shadow use is a messy signal, but it’s the only signal that tells you where the real handoffs hurt.

Given these many weaknesses, I don’t believe the headline number that 95% of company AI use failed during the first half of 2025. But I do believe that companies didn’t realize much profitability growth from AI last year. I even believe that this is still the case for most companies today. But the reason for the lack of AI profitability is the focus on task-level productivity and a refusal to redesign entire workflows for AI.

Empirical Research on Workflow Redesign

A new INSEAD/Harvard Business School paper, “Mapping AI into Production: A Field Experiment on Firm Performance,” provides causal evidence that profiting from AI is not mainly a prompting or access problem. It is a workflow-design problem: firms must discover where AI can reorganize the production system, which the authors call the mapping problem.

The “mapping problem” is the central managerial friction in AI adoption: the challenge of discovering exactly where and how artificial intelligence creates value within a production process. (NotebookLM)

I made a short explainer video summarizing the main points of this research (YouTube, 5 min.).

Conducted by researchers Hyunjin Kim, Dahyeon Kim, and Rembrand Koning, the research was structured as a massive randomized field experiment involving 515 high-growth startups participating in a global, three-month virtual accelerator program known as the AI Founder Sprint.

All 515 startups, spanning the Americas, Europe, Asia-Pacific, and the Middle East, and operating with a median team size of four, received identical, high-level access to AI technology. Every single startup in both the control and treatment groups received approximately $25,000 in in-kind API credits, access to frontier foundational models, and sophisticated partner tools from industry leaders like Google Cloud, OpenAI, NVIDIA, and Manus AI.

Furthermore, both groups received identical, intensive technical training. Every week, participants attended three-hour, hands-on sessions delivered by experts from MIT and Harvard. These sessions covered advanced AI capabilities, including rapid prototyping, retrieval-augmented generation (RAG), agentic workflows, context engineering, and “vibe-coding.” Both groups also had equal opportunities to pitch their ventures to global venture capitalists. If the barrier to profiting from AI was merely technical access or coding skills, both groups would have performed identically.

The startups were randomized into a control group and a treatment group. The single manipulated variable between the two groups was the strategic information they received during the accelerator’s weekly 60-minute workshops.

• The Control Group received standard, high-quality entrepreneurship case studies focused on established best practices. They learned about building ideal customer profiles, designing tests to validate ideas, hypothesis-driven development, and general lean startup methodologies.

• The Treatment Group received a specialized cognitive and strategic intervention explicitly designed to help them solve the Mapping Problem. In their required workshops, they were presented with case studies explicitly showing how other AI-native startups had reorganized their entire production processes around AI. They were given detailed visual diagrams contrasting “Before AI” and “After AI” workflows, illustrating how entire chains of tasks could be compressed, reordered, or completely eliminated.

A key part of this study was to show the companies concrete examples of how AI had been used for complete workflow redesign in other companies. Broad principles and theory are fine, but specifics prod people into action. (NotebookLM)

Four of the case studies were:

1. Gamma, an AI-native presentation company, compressed a multi-person, multi-month product development cycle. Traditionally, product development involves a linear chain: a product manager gathers feedback, writes a spec, hands it to a designer, who hands it to an engineer, who hands it to QA. Gamma reorganized this. AI was deployed to detect usage patterns and generate product variants directly. This enabled a single PM to continuously ship features that previously required a whole team. Crucially, the reorganization introduced a new vital task: developing “AI evals” (evaluations) to ensure the AI’s rapid generation actually improved the product.

2. Ryz Labs completely changed the prototyping phase. Conventional prototyping requires choosing one tech stack and building a single version of a product over months. With AI, the founder writes a Product Requirements Document and feeds it into multiple AI coding tools simultaneously. They build the same idea in three different ways. Then, in a single user research session, they test the best version, get feedback, use AI to rebuild the feature live on the call, and test it again instantly.

3. FazeShift, an accounts receivable startup, demonstrated how partial automation preserves bottlenecks. The conventional process involved eight steps alternating between software systems (Excel, QuickBooks, banking portals) and human clerks who acted as the “glue” bridging the systems. Instead of making the human faster at one step, FazeShift used AI to replace the human glue entirely, creating a fully automated sequence where AI pulls, reconciles, and emails, leaving humans only to resolve flagged exceptions.

4. Ranger started by selling QA testing as a manual service from day one. By doing the work manually, the founder learned exactly which steps were routine. They then used AI to automate their own internal delivery. This transformed the margins of the business from a low-margin service firm to a highly scalable product firm, pushing the need to raise outside capital to the very end of the journey, entirely changing the firm’s financial risk profile.

The Findings: Benefits from Viewing AI as a Workflow Redesign Opportunity

The results of this one change in the advice were staggering. When firms were taught how to solve the mapping problem and stopped viewing AI merely as a simple tool to speed up existing tasks and started viewing it as an architectural catalyst to redesign the firm, their performance skyrocketed across every measurable metric.

1. Expanded AI Adoption Across the Value Chain:

The treated firms discovered and implemented 44% more AI use cases than the control group (averaging 8.8 cumulative cases compared to 6.1). But more importantly, the breadth of these applications expanded significantly. Rather than just using AI for basic administrative tasks, treated firms applied AI across a wider range of distinct functional categories. They concentrated heavily on product development, strategic design, and deep business operations. They transitioned from using AI as an isolated novelty to embedding it into the fundamental mechanics and architecture of what they were building.

The best firms used AI to reorganize business operations, improve product development, and rethink their production process, sometimes for full automation. (NotebookLM)

2. Increased Output of Internal Tasks:

Treated firms completed 12% more business tasks over the course of the accelerator. To understand this, independent human coders categorized these tasks and found that the increase was driven almost entirely by internal tasks: building products, writing code, creating financial models, optimizing infrastructure, and prototyping. Because the accelerator provided identical networking and pitching opportunities to both groups, external tasks (like meeting investors or doing sales outreach) remained exactly the same. AI simply allowed the treatment group’s internal production engine to build much more in the same amount of time.

That distinction reveals a useful boundary condition: workflow redesign pays most when the firm is constrained by delivery rather than by demand generation. In this experiment, gains came from a stronger internal production engine. So the highest-return targets are businesses with real demand, real backlog, or clear product opportunity, but too much friction in fulfillment.

3. Massive Gains in Firm Performance and Revenue:

The expanded, structural use of AI aggregated into undeniable business traction and profitability. Treated firms progressed along a range of milestones much faster. They were 11 percentage points (18%) more likely to acquire paying customers. Most incredibly, they generated 90% higher total revenue than the equally funded, equally skilled control group by the end of the program.

The single most important statistic from the study: companies that were taught how to use AI for workflow redesign increased revenue by 90%. (NotebookLM)

4. The “Right-Tail” Breakout:

When analyzing profitability, it is critical to look at the distribution of success. The researchers’ quantile regression analysis revealed that the revenue gains were not distributed evenly as a modest, uniform bump for every single company in the treatment group. Instead, the gains were massive at the 90th and 95th percentiles. This indicates that reorganizing around AI doesn’t just make mediocre firms slightly better; it fundamentally expands the upper ceiling of what top-tier ideas can achieve. By removing sequential bottlenecks in the production process, AI allows the most promising ventures to break out and scale exponentially.

AI made it possible for the best ideas to grow really big. (NotebookLM)

Traditional workflows force sequential commitment: pick a stack, build for months, then learn. AI-native workflows let you hold multiple futures open until real user data closes them. That is the financial definition of a real option, and options are disproportionately valuable in the right tail. Workflow redesign doesn’t raise the average; it raises the ceiling by letting top teams run dozens of low-cost experiments where others run one. Revenue skew follows.

5. Scaling Without Proportional Inputs (Doing More With Less):

Perhaps one of the most fascinating findings relates to resource demand, operational scaling, and capital efficiency. In traditional business, if a company is completing more tasks, acquiring more customers, and generating double the revenue, they naturally require more capital and a larger headcount to sustain that growth. The treated firms broke this fundamental economic rule. Despite their massive growth, the treated firms reported needing 40% less external capital investment to reach their milestones.

Furthermore, their demand for human labor remained completely unchanged. Companies targeting workflow redesign for AI achieved hyper-growth without the traditional, dilutive costs of scaling. (Note that AI didn’t cause job losses: the workforce with and without workflow redesign was the same; increased profitability came from the same staff accomplishing more, not from firing anybody.)

Firms in the treatment group produced more but required less capital investment and used the same amount of labor. (NotebookLM)

6. Gains Are Not Technical, but Organizational:

The gains were not limited to technically trained founders, nor were they strongly concentrated among startups that were already ahead at baseline. That is good news for design leaders. It means the binding constraint was not primarily prompt engineering, coding skill, or preexisting quality. It was a cognitive and organizational search problem. The firms that won were the firms that broadened their search across the production process. In other words, product designers, service designers, and UX strategists are not secondary players in this story. They are close to the center of it, because their disciplines are about structure, flow, dependency, and use in context.

7. The Proprietary Data Flywheel:

By fully mapping and digitizing end-to-end workflows, firms unlock a massive secondary benefit: continuous, high-quality data generation. In a traditional workflow, the “human glue” holding legacy systems together leaves no digital trace of the implicit judgments employees make every day. (If a person quits, that knowledge goes poof.) Once AI acts as the continuous bridge across the entire sequence, every automated action, resolved edge case, and routing decision is captured. End-to-end workflow redesign builds the contextual data assets needed to fine-tune highly specialized, proprietary AI models.

When AI generates variants across an entire workflow, every human rejection, edit, or reroute becomes a labeled counterfactual: “here’s the output we refused, and why.” Traditional systems never capture this; the knowledge dies in Slack. Over months, firms accumulate a proprietary “rejection library” that teaches models not just your style, but your boundaries. Competitors cannot copy your taste in no’s. That is the moat workflow redesign builds, and it’s invisible until you design the workflow to log it.

What This Study Means for Companies That Want to Profit from AI

The profit implications of this research are profound. It proves empirically that the primary bottleneck to AI profitability is not technological; it is managerial, strategic, and design-oriented.

Currently, most companies view AI through a narrow lens of what strategy researchers call “local search.” Human beings possess bounded rationality. When confronted with a vast, combinatorial, and unpredictable technology like generative AI, humans default to what is familiar. If you ask a team how they can use a new technology, they will look at their current, immediate daily tasks and ask, “How can an AI do this specific thing faster?” A UI designer might use AI to generate placeholder text. A developer might use it to debug a script. A customer service rep might use it to draft a reply.

Local search is a natural instinct: look at what’s familiar, optimize where you’re at. But this is limiting when faced with a revolutionary technology like AI. (NotebookLM)

Firm production is complementary. In a tightly coupled sequence, speeding up one step rarely raises firm-level throughput; it just moves the bottleneck downstream. Make UI design 300% faster, and the front-end backlog gets longer. Partial automation preserves bottlenecks.

To truly profit from AI, companies must deploy their product designers, service designers, and UX strategists, to map the entire production process. Profitability requires cross-functional reorganization. It requires viewing AI not as a tool for a specific worker, but as an opportunity to collapse entire sequences of activities, bypass traditional resource constraints, and redefine the business model itself. The control group in the experiment had the exact same technology, but they failed to profit from it because they didn’t know how to map it. The companies that win the AI revolution will be those whose design professionals successfully blueprint these new, AI-native architectures.

Across the cases in the study, four redesign moves recur. Remove handoffs. Parallelize variants. Move humans to exceptions. Add evaluation loops. That checklist is more actionable than the vague instruction to “use AI more.” If a proposed AI initiative does not do at least one of these four things, it is probably only a local optimization.

Recommendation 1: Redesign Service Blueprints to Eradicate “Human Glue”

Action Item for Service Designers and Operations Strategists.

Service design has traditionally focused on mapping the end-to-end customer journey, documenting both the front-stage (what the customer sees) and the back-stage (the internal employee processes that deliver the service). In many organizations, especially in B2B, healthcare, finance, and operations-heavy environments, the back-stage is held together by what we can call “human glue.”

As demonstrated in the study’s FazeShift case study, human workers constantly act as manual bridges between disconnected, legacy software systems. For instance, a traditional accounts receivable process involves an employee downloading data from an Excel spreadsheet, manually entering that data into QuickBooks, checking a banking portal to match payments, and then typing a confirmation email in Gmail. When companies first adopt AI, they attempt to augment this flow. They might give the clerk an AI writing assistant to draft the email faster. But the research proves that applying AI to augment just one of these nodes yields negligible firm-level returns. The bottleneck is the human bridging the gaps between the software; the speed limit of the firm is still dictated by manual data entry.

Humans bridging gaps become a limiting factor. (NotebookLM)

The job for service designers is to actively hunt for, map, and design out this human glue. In a service blueprint, mark every moment when a person manually transfers context, data, or state between systems. Then redesign the sequence so AI becomes the bridge and the human moves to the exception path: auditing outputs, resolving flagged cases, and setting thresholds. That is the shift from node-level augmentation to end-to-end automation.

In this redesigned blueprint, the human’s role is elevated entirely out of the routine sequence. The human is placed at the very end of the pipeline solely to handle edge cases, audit the system, and resolve flagged exceptions that fall below a specific AI confidence threshold. By designing full-chain automation rather than node-specific augmentation, service designers can help their companies turn expensive, labor-intensive professional services into highly scalable, high-margin product offerings. This is how you unlock the massive 90% revenue lift seen in the study while keeping labor costs flat.

Recommendation 2: Transition from Linear Agile to Parallel AI Prototyping

Action Item for Product Designers and Product Managers.

Historically, digital product design has been a highly sequential, highly specialized discipline. The traditional Agile flow (which the research explicitly identifies as a legacy “Before AI” bottleneck) begins with user research, moves to a product manager for feature prioritization, flows to a UX/UI designer for wireframing and high-fidelity mockups, hands off to an engineering team to build a single version, passes to QA for testing, and finally reaches a data scientist for A/B evaluation. This process is slow, expensive, and forces teams to make high-stakes bets on a single design concept because the cost of building multiple, fully functional coded prototypes is economically prohibitive.

In a sequential process, everybody is lined up, waiting for the previous step to finish before they start working. This stretches out the total time until you have the result. (NotebookLM)

Based on the research findings, specifically highlighted in the study’s Ryz Labs case study on parallel prototyping, product designers should stop acting as a specialized node in a long, linear chain. As AI coding and interface-generation tools drive the cost of working software toward zero, designers become parallel generalists who can explore several build directions at once.

In parallel prototyping, you build many varying solutions simultaneously, so that you can test them before proceeding to make the final, shipping version. This would be prohibitively expensive with legacy manual design and implementation, but when AI does the work, you just let multiple agents lose on the project. (NotebookLM)

When a new feature is conceived, do not spend weeks perfecting one static design to hand off. Use AI to generate several working prototypes from the same PRD, then test, revise, and retest them during the same research session. The designer’s role shifts from authoring a single artifact to orchestrating multiple software variants.

Recommendation 3: Embed AI Evaluations as a Core Design Discipline

Action Item for UX Strategists and UX Researchers.

One of the most vital insights from the experiment’s exploration of AI-native production (highlighted in the Gamma case study) is that when AI is used to drastically increase the speed and volume of product generation, the organizational bottleneck immediately shifts from creation to evaluation. If an AI system can instantly detect user behavior patterns and autonomously generate personalized interface variants to solve user drop-off, the limiting factor is no longer how fast we can build the UI. The new limiting factor is determining whether the AI-generated UI is actually good, safe, on-brand, accessible, and effective.

To address this new bottleneck, UX teams must pioneer and standardize an entirely new internal discipline: “AI Evals” (Evaluations). Alongside creating design systems and component libraries, UX designers must create robust, automated evaluation criteria that judge the quality of AI outputs at massive scale. Designing an “eval” requires codifying human UX intuition into programmatic, machine-readable rules. What makes an interface accessible? What constitutes off-brand language? What interaction patterns are considered unethical dark patterns?

Quality control at scale cannot remain a manual process. We need AI to conduct our AI evals. (NotebookLM)

Designers need guidelines, prompt structures, and heuristics that let a second AI system grade the first AI before it reaches users. Researchers, meanwhile, should shift from testing static wireframes to monitoring live, AI-generated behavior in production. UX professionals must evolve from makers of artifacts to judges of algorithmic outputs. If you do not speed evaluation along with generation, you scale bad experiences faster.

Recommendation 4: Invert the Business Model from Product-First to Service-to-Product

Action Item for UX Strategists, Service Designers, and Innovation Leaders.

The study highlights a profound finding regarding resource allocation: treated firms required 40% less external capital to achieve superior growth. This wasn’t merely because they saved money on software subscriptions; it was because solving the mapping problem allowed them to fundamentally re-sequence how they built their business models. The traditional product development path is highly front-loaded: raise millions of dollars, hire a massive design and engineering team, spend a year building a SaaS product in a vacuum, and then attempt to sell it to customers. This requires massive upfront capital and carries immense product-market fit risk.

UX strategists, especially those working in startups, corporate innovation labs, or new venture incubators, should flip this sequence using a Services-First approach, as illustrated by the study’s Ranger case study. Instead of designing a massive software product from day one, strategists can design a high-touch, human-delivered service. The founders, designers, and core team deliver the value manually to the customer, operating as a concierge or consulting service. This allows the team to intimately learn the customer’s exact pain points and establish deep tacit knowledge of the operational process.

Because AI drastically lowers the cost of coding and workflow automation, the team can use the revenue and deep insights generated from the manual service to iteratively build bespoke AI tools that automate their own internal delivery. Over time, the internal tools become so robust that the margins shift from those of a traditional service agency to those of a highly scalable software company. By starting as a service and letting AI lower the costs of delivery with each cycle, UX strategists can validate product-market fit on day one, generate immediate revenue, and delay (or completely eliminate) the need to raise dilutive external capital. You are effectively getting paid to do user research while building the product in the background.

Recommendation 5: Break Local Search Bias with “Before and After” Visual Frameworks

Action Item for Design Leadership, Design Ops, and UX Directors.

Perhaps the most fundamental, meta-level finding of the research is that human beings are naturally constrained by local search. When faced with a vast, unpredictable, and combinatorial technology like AI, people default to what is familiar. If you ask a marketing team to adopt AI, they will use it to write copy. If you ask an HR team, they will use it to summarize resumes. This is the very essence of the “Mapping Problem,” where teams fail to see the transformational opportunities because they only look at the steps immediately in front of them. The treatment in the INSEAD/HBS experiment worked so effectively because it provided founders with specific, structured visual frameworks that forced them to abstract their thinking and look at the entire firm’s architecture.

Design leaders must actively disrupt local search within their teams and organizations. Do not simply hand your designers a premium AI subscription and ask them to “be more productive.” Instead, facilitate mandatory, cross-functional mapping workshops. Design Ops leaders should enforce the use of service blueprints and value stream maps that visualize the entire lifecycle of a product or service, from the moment a customer intent is formed to the final reconciliation of revenue.

During these workshops, leaders must introduce explicit “After AI” visual frameworks (exactly like the diagrams used in the experiment) to challenge entrenched assumptions. Put the current process on the whiteboard and force teams to answer provocative, constraint-based questions: “If we had to deliver this feature in 24 hours without any human engineers, how would AI do it?” or “If we could eliminate three handoffs in this specific process, where would an AI agent need to sit?” By institutionalizing the use of systemic mapping frameworks and providing “discovery scaffolding,” design leaders can elevate their teams from local task optimization to holistic, high-value process transformation. This structured exploration is the exact mechanism that led treated firms to discover 44% more AI use cases and achieve broad-based innovation.

Clearly visualized before-and-after scenarios show the effect of simplifying workflows with AI. (NotebookLM)

Recommendation 6. Replan Roles, Hiring, and Capital Needs Around AI Leverage, and Make That Replanning Continuous

Action Item for C-Level Executives.

Many firms still estimate headcount and funding needs as if work will continue to flow through yesterday’s production model. Once AI changes the model, those estimates can become badly inflated. For UX strategists, this is a reminder that operating assumptions are part of the design brief. For product and service leaders, it means the team model itself may need redesign: fewer people doing repetitive glue work, more people orchestrating systems, defining policies, reviewing exceptions, and improving evaluations.

The action is to revisit staffing and financing logic whenever a meaningful AI-enabled redesign lands. Instead of asking, “Which role do we hire next?” ask, “Which bottleneck still requires human judgment, and which role best improves that judgment?”

At the same time, do not treat AI mapping as a one-off transformation workshop. The paper ends with a critical observation: as models improve, the search space gets larger, which means the mapping problem may get harder, not easier. So build a cadence. Review newly discovered use cases monthly or quarterly. Keep an internal case library. Record what changed in the workflow, what metric moved, what new bottleneck emerged, and what role definition changed with it. The firms in the study benefited from repeated exposure to examples and structured thinking, not from a single inspirational moment. Your company needs the same discipline if it wants durable returns rather than a short-lived burst of AI enthusiasm.

One obstacle is political, not just cognitive. Every handoff usually belongs to a team, a budget, and a manager, so workflow redesign threatens local turf even when it improves firm-level performance. That is why leaders need shared success metrics and executive sponsorship. Otherwise, each function will optimize its own step, and the bottleneck will simply reappear somewhere else.

Recommendation 7: Design the Employee Experience (EX) for AI Orchestration

Action Item for UX Strategists and HR Leaders.

When you eradicate the “human glue” and elevate employees solely to handle edge cases and audit AI outputs, you fundamentally alter their daily cognitive load. Transitioning from a “doer” of routine tasks to an evaluator of complex machine-generated output can lead to cognitive fatigue and automation bias (the human tendency to blindly trust the machine). UX professionals must apply user-centered design to the employees’ new reality. This means designing internal exception-handling interfaces that go beyond simple “approve/reject” buttons to provide transparent algorithmic reasoning and meaningful context. Redesigning the workflow is only half the battle; you must also intentionally design the daily digital experience of the humans overseeing it.

Automation bias is the tendency to follow where AI leads, especially when AI has been proven right most of the time in the past. We need new design principles that prevent users from approving AI when it’s wrong, without lulling them into complacency by asking them to approve myriads of cases when it’s right. (NotebookLM)

The Unproven Frontier

My seven recommendations in the previous sections are directly based on the 2026 INSEAD/Harvard field experiment. However, for product designers, service designers, and UX strategists to maintain a durable competitive edge, they must also look past the immediate horizon and consider radical future changes. The following four futuristic ideas are my challenge to you for possible futures to consider if you want to push yourself.

Let’s look past the current findings to speculate about frontier use of AI over the next several years. Be warned that not all these ideas may work out. (NotebookLM)

Speculation 1: Agent-to-Agent UX Design (Zero-UI)

Historically, the end-user of any digital product or service has been a human being. The entire discipline of user experience is built around concepts from human psychology, such as visual hierarchy, Fitts’s Law, and cognitive load. However, as autonomous AI agents become ubiquitous, we are entering a reality where the primary “user” of your software will be another piece of software acting on a human’s behalf.

If a consumer uses a personal AI agent to scour the internet, evaluate prices, and purchase auto insurance, how does an insurance company design its digital presence to “convert” that AI agent? You aren’t trying to appeal to human emotion, color theory, or clever copywriting. You are designing for machine readability, structured data, API robustness, and logical consistency. UX strategists will need to invent the field of “Agent-to-Agent Experience.” (Maybe we’ll call this AX instead of UX.)

As AI agents take over more of the work, human users will step back and often don’t interact with your website or service. Instead, agents will be interacting with other agents, in an AX (agent-to-agent experience).

This involves designing parallel experiences: a human-facing experience that emphasizes trust, and an agent-facing experience optimized for extremely low-latency data parsing. The companies that figure out how to make their products frictionless for AI agents to discover, evaluate, and purchase will capture a massive share of the emerging machine-to-machine economy.

Speculation 2: Self-Healing and Auto-Optimizing Service Blueprints

The INSEAD/Harvard study demonstrated the value of mapping out the human glue and manually designing an AI sequence to replace it. My speculative extension is that the service blueprint itself will cease to be a static document created by a human designer. Instead, it will become a living, “self-healing” software entity deeply integrated into the company’s operational tech stack.

Today, most service blueprints are static workshop artifacts. In an AI-native company, that may start to look absurd. A more advanced practice would turn the blueprint into a living operational map connected to real telemetry: time spent at each step, exception rates, escalation frequency, satisfaction drops, model-confidence distributions, and unit-cost movement. The blueprint would stop being a document and become a diagnostic surface. Design and operations leaders could then see where AI has actually shifted the bottleneck this week, not where they think it should have shifted based on last quarter’s workshop. That could make service design much more like continuous product analytics.

In this future, an overarching systemic AI will constantly monitor the flow of data, capital, error rates, and customer satisfaction across the entire value chain of the firm. If a bottleneck emerges, say, a sudden influx of customer support tickets regarding a specific billing error or a confusing UI flow, the living blueprint will detect the friction autonomously. It will then automatically write a script to patch the billing software, update the customer-facing FAQ, generate a clearer UI modal, and instantly reroute human exception-handlers to manage the overflow, all while updating the visual blueprint dashboard for human managers to review.

Service designers will program the parameters, ethical boundaries, and fail-safes of an autonomous, self-optimizing organizational nervous system.

Speculation 3: Workflow Architect as a New Cross-Functional Role

In the study, the firms benefiting from AI were not merely better operators in the classic sense. They were better at seeing how multiple activities fit together and how those activities should change together. In larger organizations, that capability may become too important to leave scattered across job descriptions.

A plausible next practice is the emergence of a dedicated role that sits between product, service design, operations, and data: call it a workflow architect, AI operating designer, or production strategist. The title matters less than the mandate. This person would map dependencies, identify where AI reduces one constraint and creates another, define exception logic, and coordinate redesign across teams that would otherwise optimize locally.

That role does not appear in the paper, but the need for it follows from the paper’s argument. The more capable AI becomes, the wider the search space becomes, and the more expensive local thinking becomes. A workflow architect could institutionalize the very thing treatment provided in the experiment: structured, cross-boundary reasoning about where AI belongs and what surrounding changes are required. In companies with complex services, that could become a major source of competitive advantage.

The workflow architect is a cross-functional role. You need the ability to span two domains. (NotebookLM)

Speculation 4: The Rise of the Solo-Decacorn and Zero-Capital Corporate Venturing

The research highlighted a stunning 40% drop in the capital required by startups to reach their milestones when they mapped AI effectively. If we extrapolate this finding to its absolute limit, we can speculate a massive revolution in corporate venture creation and entrepreneurship. Currently, large companies struggle to innovate because launching a new digital service requires budgeting millions of dollars for cross-functional teams, navigating immense bureaucratic friction, and waiting 18 months for delivery.

The need for capital investment drops with AI, vastly expanding the opportunities for entrepreneurship. If you have an idea, go for it, without needing to convince a VC. (NotebookLM)

In the future, a single UX strategist and a product manager inside a Fortune 500 company (or operating independently in a garage) could act as a fully self-contained venture studio. Armed with a suite of AI agents that map existing data lakes, automate the coding process, and run synthetic marketing tests, this two-person pod could spin up fully functional, enterprise-grade new services with effectively zero budget. We may see an era of “Zero-Capital Corporate Venturing,” where the cost of internal experimentation drops so low that corporations can afford to launch, test, and kill hundreds of fully realized digital products a month. Furthermore, we will likely see the first “Solo-Decacorn” in the form of a ten-billion-dollar valuation company operated by a single designer who masterfully mapped an entire corporate architecture using AI agents.

A single person creating $10 billion in value. Would have sounded crazy in the pre-AI age, when even the most brilliant founder needed employees to scale a business to that volume. But with AI doing the work, we will likely see solo-decacorns within the decade. There is already at least one solo-unicorn (a company doing $1 billion in annual revenue). (NotebookLM)

Past Workflow Redesign to Full Company Redesign

What’s the next step after redesigning end-to-end workflows for AI? The logical conclusion is designing the entire company for AI. If workflow redesign strips away bottlenecks to execute the current business faster and cheaper, company redesign means architecting the organization to do things that were previously physically or economically impossible.

AI necessitates new company architectures that are different from the traditional ones. (NotebookLM)

For more than a century, firms were built around the scarcity of human cognition and the high cost of coordination. Specialists were expensive. Translation across functions was lossy. Reporting was manual. Regulatory interpretation was slow. International expansion required local staff. That is why companies accumulated departments, regional silos, thick handoffs, and multiple layers of management. Those structures were not timeless truths. They were coping mechanisms for a high-friction economy.

The pre-AI world was filled with friction, making it hard to do business. AI greases the wheels of commerce. (NotebookLM)

AI lowers the friction. When language, coding, summarization, translation, analysis, and first-pass compliance become dramatically cheaper, the optimal shape of the firm changes with them. The important question is no longer only, “How do we redesign this workflow?” It becomes, “If we were founding this company today with AI as a native capability, would we organize it this way at all?”

AI allows us new ways to organize the firm, opening up new possibilities for where to go. (NotebookLM)

That is a deeper reconceptualization than automation. Workflow redesign removes bottlenecks inside an existing sequence. Company redesign questions which sequences should exist, which departments should exist, which customers should be served, and which offerings become economically possible. AI does not merely make the old firm run faster. It expands the possibility frontier of what a firm can be.

For UX strategists, design leaders, and business founders, this reconceptualization will manifest across four radical shifts in corporate architecture:

1. “Impossible” Products and Mass-Bespoke Economics

Currently, businesses are constrained by the economics of human labor: you must choose between offering cheap, standardized products to the masses (like a one-size-fits-all SaaS platform) or expensive, bespoke services to a few (like a boutique consulting firm). AI breaks this fundamental economic tradeoff. When the marginal cost of expert cognitive labor and custom code drops to near-zero, companies can launch completely new products that were previously financially unviable.

Imagine a software company that no longer sells a rigid platform, but instead deploys an AI system that analyzes a client’s specific operational quirks and dynamically generates a bespoke, one-of-a-kind platform entirely from scratch in an hour. Mass-market “Service-as-AI-Software” (the new SaAS acronym) will become the new standard. Service designers will no longer design a single static service; they will design the systemic parameters and ethical guardrails within which AI can invent hyper-personalized services for a “segment of one.”

AI enables bespoke products and services at scale. Everybody can get exactly what they want, even from a small vendor. (NotebookLM)

2. The “Micro-Multinational” and Frictionless Global Reach

Historically, expanding to reach new global customers required massive capital to establish local headquarters, hire regional compliance officers, and deploy localization teams. Large enterprises have long used these complex legal and linguistic barriers as a defensive moat against smaller competitors. AI evaporates this moat entirely.

Through real-time, context-aware AI translation and automated regulatory compliance, a three-person startup or a tiny boutique vendor will be able to distribute complex services worldwide on day one. Operating as a “Micro-Multinational,” their AI agents won’t just translate the text; they will dynamically adapt the UI for cultural resonance, and autonomously rewrite legal contracts, payment routing, and onboarding flows to comply with local privacy laws and trade tariffs. The strategic design challenge shifts to creating a universal base platform that AI can safely and legally adapt to infinite global contexts on the fly.

AI will give tiny organizations worldwide reach. (NotebookLM)

3. Morphing Company Boundaries

AI also reopens the question of firm boundaries. Some activities that once had to be internal because coordination was too costly can now be orchestrated externally through AI. (This is sometimes called “The Coasean Singularity.”) Other firms will pull more variation inward because AI makes internal recombination cheap. The point is not that every company becomes smaller or larger. The point is that the boundary between the firm and the market becomes a design choice again.

What’s inside the company walls, and what’s outside? Those decisions will change with AI. (NotebookLM)

The organizational implications are equally radical. If AI can handle much of the routing, summarizing, monitoring, reporting, scheduling, drafting, and cross-functional translation that once absorbed so much managerial effort, then the classic corporate pyramid becomes harder to justify. Much of twentieth-century middle management existed to gather information, compress it, pass it upward, translate it into tasks, and chase status across departments. That was rational when coordination was expensive. It becomes far less rational when coordination is cheap.

4. The Pancaked Organization and the End of Middle Management

The legacy corporate pyramid was built to solve an information-routing problem. Middle managers historically existed to act as “human APIs,” aggregating data from the bottom, passing it up to executives for decision-making, and translating those decisions back into discrete tasks for the frontline. In an AI-native company, intelligent agents handle this data synthesis, tracking, and cross-departmental coordination instantaneously.

Consequently, the corporate structure will “pancake”: flattening from a sprawling hierarchy into a radically decentralized, ultra-thin layer. Middle management will effectively dissolve, replaced by tiny, highly autonomous execution pods. A tight-knit team of three or four human orchestrators, heavily leveraged by swarms of AI workflow agents, will yield the output, strategy, and operational effectiveness of a traditional 300-person department from the “Before AI” world. Bye-bye, VPs.

AI will enable pancaking, where high-value companies (or focused teams within larger companies) are led by a single visionary founder supported by a handful of high-agency employees, without middle management to delay execution. Crash go the big edifices constructed by traditional hierarchical management, as well as the career ladder some people have been planning to climb. (NotebookLM)

In this pancaked reality, architecting the company structure merges with service design. The ultimate deliverable of the future UX leader will not just be a product interface; it will be the operational blueprint of the company itself.

Workflow redesign was the opening move. The deeper opportunity is company redesign. New offerings become feasible. New customers become reachable. New geographies become economical. Old org charts become historical artifacts. The winners will not be the companies that use AI to run yesterday’s business slightly faster. The winners will be the companies that ask a much more dangerous question: if intelligence, coordination, translation, and adaptation are now cheap, what should this company look like at all?

One conceptual sentence sits underneath this whole section: AI is not only a labor-saving technology; it is a coordination technology. That is the reason the design target expands from task, to workflow, to firm.

Finally, AI Consulting that Works

I used to be very negative about the future of consulting and design agencies in the AI era, for the simple reason that consultants rely on established best practices to produce specific recommendations for clients. (A tiny fraction of consultants may be geniuses who can invent something completely new that’s independent of best practices to utterly redefine a client’s business. But that’s the exception, not the rule.)

The shield wall was a best practice for Viking battles. Unfortunately, best practices for AI haven’t yet been established. (NotebookLM)

Best practices work brilliantly for web usability. We’ve known since the late 1990s how people use websites and what it takes to design a good site. In fact, I discovered several of the most basic web usability guidelines in about a week, during my very first web user research at Sun Microsystems in 1994.

Unfortunately, nobody knows how to use AI to improve company profitability. This makes even the most highly paid “strategy” consultant virtually worthless, despite their usual fees of a million dollars or more per project.

AI consulting has been useless because nobody knows how to most profitably use AI in companies. We’re in unknown territory, where the most high-paid consultants don’t know anything. (NotebookLM)

Worse, the specifics of how to use AI most profitably change every year as more capable models emerge.

But this research does reveal one practice worth teaching: stop chasing isolated task productivity and redesign end-to-end workflows around AI. As more case studies accumulate, consulting becomes useful again to the extent that it helps firms map those patterns onto their own operations.

We are finally getting practical examples of companies that use AI well. (NotebookLM)

While AI consultants still don’t have the answers to what will work the best in your company, they can at least now start teaching principles and case studies to drive the redesign of how you do business. (NotebookLM)

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,050 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: OpenAI to emphasize usability | Integrated software | Comparing multi-shot AI video models | Grok’s image generation model receives an upgrade | AI use up in the workplace, and employees who use AI now worry less about losing their jobs to AI

UX Roundup for April 6, 2026 (Nano Banana 2)

OpenAI Wants AI Usability

OpenAI’s press release announcing a $122 billion funding round last week (at a valuation of $852 billion) contained an interesting statement: “As models become more capable, the limiting factor shifts from intelligence to usability.” As an old usability freak, I can only approve of this insight, late as it may be.

I would have advised OpenAI several years ago to emphasize the usability of AI much more than they have done traditionally. Maybe they could have avoided the Sora debacle if they had conducted more up-front research on user needs in AI video.

OpenAI will supposedly prioritize usability. Let’s hope so. (Nano Banana Pro)

The press release states that OpenAI’s new strategy is to build a unified AI superapp. (Maybe along the lines of the pre-AI WeChat superapp that’s been so successful in China.) They claim that “Users do not want disconnected tools. They want a single system that can understand intent, take action, and operate across applications, data, and workflows.” I have not seen user research to support this claim, but the WeChat experience does indicate the benefits of a unified superapp relative to the fragmented app experience on mobile devices outside China.

In general, integration often benefits user experience. If nothing else, a single integrated UI usually has better consistency, unless it’s produced by truly incompetent designers who make each segment disconnected from the rest. An integrated system also offers a more seamless user experience, without the clunky handoffs that come with combining disparate applications.

Historically, integrated systems succeed by drastically reducing interaction cost. Before Microsoft Office conquered the market, users juggled WordPerfect, Lotus, and Harvard Graphics, each inflicting idiosyncratic menus and proprietary formats. Combining these tools eliminated the friction of context switching. Users learned a single mental model, and data interoperability became seamless, allowing users to embed a financial chart into a document without jumping through complex export-import hoops.

An integrated user experience can support a smoother workflow. (Nano Banana Pro)

These exact historical benefits apply perfectly to the current, highly fragmented AI landscape. Today, using generative AI for a multi-step project is a miserable experience reminiscent of 1980s software. If a user wants to analyze a dataset, generate a chart, and draft an executive summary, they often have to bounce between three different tools. Each AI requires its own distinct prompting style and has zero memory of what happened in the other window. The user is forced to endlessly copy and paste outputs, essentially acting as a manual bridge to connect isolated models. This constant articulation work kills productivity and alienates mainstream users who do not have the patience to become expert “prompt engineers.” (Nor should they have to, as I’ve argued for years.)

If OpenAI actually builds a properly integrated superapp, these legacy usability victories will transfer beautifully to the AI era. The most crucial advantage of a unified AI system isn’t just a shared graphical interface, but a shared, persistent context. In such a system, the AI maintains a continuous understanding of the user’s overall intent, project history, and personal preferences across every task. A user wouldn’t have to write a three-paragraph prompt explaining their goal from scratch every time they switch from text generation to data analysis or image creation. The system already knows what they are trying to achieve.

But persistent context also creates a new requirement: inspectability. Users will not trust a superapp merely because it is convenient. They need to see what sources the AI used, what assumptions it carried forward from earlier tasks, and how to correct or reset that context when it drifts. A unified system that is opaque can be more dangerous than a fragmented system that is merely inconvenient.

Current AI tools lack the benefit of classic integrated software — and elephants — of having a persistent context that transfers from one step to the next, even when using different AI models. (Nano Banana Pro)

Ultimately, an integrated AI ecosystem can finally deliver on the holy grail of AI usability: shifting the burden of work from the human back to the computer. If the AI inherently understands the broader goal, it can draft the text, pull the data, and orchestrate the workflow in one fluid environment. If OpenAI’s designers can pull this off without succumbing to the massive feature bloat that typically plagues superapps, their new focus on practical usability over raw algorithmic intelligence might just be their smartest move yet.

Integrated Software

In 1985, I worked at the IBM T.J. Watson Research Center, which at the time was one of the world’s 3 top user interface research labs. My main project that year (together with Robert L. Mack and others) was a field study of how business professionals used integrated software, which was a new innovation at the time.

You can download the full paper about this study. (To its credit, IBM Research maintains an excellent archive of its old papers.)

We identified 6 types of integration in 1985:

• Application integration (easy access to multiple applications, using results of one program in another, such as using numbers from a spreadsheet to build a business graph).

• Media integration (working with multiple media formats together).

• Interface integration (uniform interaction with the different parts of the system).

• Systems integration (making use of multiple computer systems working together).

• Documentation integration (the system itself helps you use it, by making user assistance part of the user experience).

• Outside integration (interfacing with the “real world” outside the computer).

The 6 forms of integration. (NotebookLM)

Here’s a summary of what we discovered about how these forms of integration were used in the workplace. These conclusions remain relevant to OpenAI’s project and other efforts to build integrated AI.

Our field study of business professionals, based on 54 surveys and 12 in-depth work analysis interviews, found that integrated software was underutilized because software vendors focused on the wrong problem. They optimized product integration (making tools fit together) when users actually needed task integration: support for the work they were trying to do.

Product integration is a design strategy that makes all the basic tools fit together. Here, helmet, sword, and shield all in one. Conversely, task integration designs an integrated user experience based on what users want to do. Here, not just raiding the English, but bringing the loot home. (NotebookLM)

The Heterogeneous Reality of the Workplace

Software designers often build for a mythical, uniform user base operating in pristine environments. The real world is messy and wildly heterogeneous. We found professionals juggling an average of five different software packages, often cherry-picking standalone tools rather than relying on bloated integrated suites.

Designers may dream of a uniform user base, but in reality, people differ dramatically in needs and skills. (NotebookLM)

Furthermore, user expertise varies drastically. While every office has a few “gurus” who eagerly dive into complex systems, the vast majority are discretionary users. They want to know as little about the computer as possible. If a system lacks approachability or requires reading a massive, intimidating manual, these professionals will exercise their veto power. They simply refuse to use it, either abandoning the task entirely or delegating it to specialized support staff.

Many teams and workplaces have a “guru” user, who may have a regular job title, such as lawyer or Viking warrior, but who loves learning about the newest technology and adapting it to local use. Most employees are not like that! (NotebookLM)

The design implication for AI is straightforward: build for discretionary users first and experts second. Give beginners a forgiving default path that works with natural language and minimal setup. Then let power users expose deeper controls (reusable workflows, reference assets, parameter settings, and debugging views) only when they actually need them. Progressive disclosure beats feature dumping.

The Power of “Satisficing”

When users do interact with software, they rarely seek the optimal technical solution. Instead, they rely on a behavioral strategy called satisficing: finding a method that is just “good enough” to handle the immediate problem.

Satisficing: once you find a satisfactory solution that’s good enough, people tend to keep using this approach even if it’s suboptimal. Sticking with what you know is easier than finding something new. (NotebookLM)

For example, we frequently observed users forcing complex data into simple spreadsheets instead of using a proper relational database. Is a spreadsheet the right tool for complex queries? No. But from a usability perspective, it is a highly rational coping mechanism. The spreadsheet is familiar, and the cognitive overhead required to learn a new system is perceived as an insurmountable barrier.

This creates a secondary usability disaster: the migration trap. Users predictably outgrow their initial ad hoc solutions as data accumulates, yet they remain trapped. The industry provides almost no graceful migration paths from introductory interfaces to mature, advanced tools.

A modern AI product should therefore support graduated migration. A novice may begin with one-off chats, then save a good interaction as a template, then turn that template into a reusable workflow, and only later into a full agent. If every increase in power requires users to abandon the old interface and learn a new one from scratch, most of them will never make the climb.

Coping With a Non-Ideal World

System designers build software for an idealized environment. In reality, users were constantly thwarted by absurdly trivial limitations: missing cables, lack of memory, or small screens that cause a frustrating “workspace squeeze.” Users also suffer from fundamental interface flaws, with over half complaining about low-level inconsistencies like function keys arbitrarily changing meaning between applications. Software must be designed for graceful degradation, delivering productivity gains even when operating in less-than-perfect, sub-optimal configurations.

While many of the specific problems that plagued our users 41 years ago have (slowly) been solved by industry standards such as universal USB-C cables, new practical problems invariably rise to take their place.

The Catastrophic Overhead of “Small Things”

Perhaps the most glaring usability failure of integrated software was its inability to handle the micro-tasks of daily office work.

Professionals constantly need to move tiny pieces of information, like pulling a single statistical number from an analysis tool into a report. The computer-based integration in 1985 was horribly heavy for this. The procedural interaction cost of selecting, copying to a clipboard, switching applications, and pasting was simply too high.

Consequently, for integrating many “small things,” the most efficient tool was still a physical piece of paper. The user writes the number down on a notepad and retypes it. When a piece of paper beats your computer system, your user interface has failed. Paper offers absolute flexibility and near-zero cognitive overhead. Software must mimic this frictionless data transfer.

This matters even more for AI than it did for office suites. The biggest gains will often come not from spectacular end-to-end autonomy, but from eliminating dozens of tiny frictions: carrying one number into a slide, one decision into a task list, or one instruction from yesterday into today’s work without making the user repeat it.

The Usability Bottom Line

To design effective tools for knowledge workers, we must stop obsessing over product integration: making proprietary code modules talk to each other. Users do not care about your software’s technical architecture. We must focus on task integration, minimizing the mismatch between the system and the user’s actual workflow. Design for the busy, pragmatic professional whose primary goal is getting the job done with minimum effort.

Comparing Multi-Shot AI Video Models: Seedance 2 vs. Kling 3

I used two different video models to make the same test scenes from Shakespeare’s “Richard III”: Seedance 2 and Kling 3. These are currently considered the world’s best AI video models. Both are Chinese, and both offer multi-shot sequences, where the AI composes a longer segment from shorter clips with different camera angles.

I set up a video model battle comparing Seedance 2 and Kling 3, having both generate the same scenes from Shakespeare. (Nano Banana 2)

I just used Seedance 2 with great success for my music video based on Hamlet (YouTube, 4 min.), but I used the omni-reference ability of Seedance for this video, meaning that I gave it character reference sheets to show how each character should look, as well as style reference images to lock in the 3D CGI style I had chosen for this project. I even gave Seedance specific renders for each scene to show it how I envisioned the action in each scene.

I took a stricter approach for this project. To see how these models perform unassisted, this Richard III experiment relied purely on text-to-video generation. I provided only scene descriptions, dialogue, and brief outfit notes, forcing Seedance and Kling to generate the visuals, character voices, and lip-syncing entirely from scratch.

Result: This small experiment clearly showed why most creators say that Seedance is the world’s best video model. (We can hope that Google will catch up with Veo 4, whenever that’s released.)

Kling 3 produced attractive scenes, but Seedance 2 won on cinematography, acting, and set design, especially in the Tower of London segment, where Clarence is being detained for his own “protection.” Seedance also delivered slightly better lip sync and fewer pronunciation errors.

Both models mispronounced some of Shakespeare’s lines, particularly archaic English, such as the famous opening line, “Now is the winter of our discontent.” Kling 3 completely garbled “discontent,” whereas Seedance 2 pronounced it “discomfort.”

Since both video models are Chinese, it may not be surprising that they have trouble with rare English words. By contrast, ElevenLabs pronounced “discontent” correctly in the closing avatar narration, suggesting that video generation is advancing faster than speech quality in these end-to-end systems.

As video models improve their visual and cinematic capabilities, sound is lagging behind, especially in foreign languages. (Nano Banana Pro)

Ultimately, “best video model” will not be judged on a single dimension. Creators care about cinematic quality, character consistency, prompt adherence, editability, speech intelligibility, rendering speed, and cost. Seedance may be the clear winner in this Shakespeare test, but another model could still be the better production choice for ad work, fast iteration, or tightly controlled brand content.

Grok Imagine Update

Grok has released a new version of its Imagine image and video model. The main touted improvement is higher image quality. I still rank Nano Banana 2 higher overall, but Grok is improving at an impressively high rate, with the next release already in the works and said to be only a few weeks away.

It’s always great to see competition, and there are already use cases where Grok Imagine is superior because it’s less censored than Google.

User testing as a modern video game. (Grok Imagine)

User testing as a classic video game. Note how Grok included a SUS score (SUS = System Usability Scale). This demonstrates that Grok’s image generation is integrated with the main AI model and its world knowledge. (Grok Imagine)

Increased AI Use in the United States

This is not a big surprise, but the latest consumer survey shows that Americans are using AI more. TD Bank surveyed 2,504 Americans in February 2026 and found that 78% of Americans now report using AI-enabled tools in their daily lives, with two-thirds saying they are more proficient than they were last year (67%).

Use is even higher in the workplace, with 83% of employed respondents indicating they use AI-powered tools or applications to support their work. 68% of the people using AI for work believe AI makes them more productive in the workplace.

AI is changing from a novelty to a normal, expected core component of the workplace, according to the new survey. (NotebookLM)

Another notable result is trust. AI is now trusted “a great deal” or “somewhat” by 62% of Americans, up from 50% a year earlier. That puts it ahead of news media at 58%, though still far below “Your Bank” at 85%. Trust is improving, but AI has not yet achieved institutional-level credibility.

Shadow AI is still real: In the same survey conducted in March 2025, 41% said “I use AI at work without my boss knowing about it.” Worryingly, 38% said that they have “exaggerated my ability to use AI when applying for my job.” Hiring managers beware! When hiring new staff, you need to give them a test to see how strong their AI skills actually are.

Whereas in last year’s survey, many people worried about AI taking their jobs, this year, 68% of survey respondents who use AI to assist with work stated they are less worried about AI taking their jobs than they were a year ago. 71% said that AI gives them an advantage over others in similar roles. (NotebookLM)

The most common AI uses at work in the 2025 survey were writing or editing (46%), conducting research (37%), and analyzing data (35%). The least common uses were recruitment (11%), optimizing supply chains (13%), and developing products and services (16%). I would say that AI is great at helping with product and service design, but maybe the low score reflects how few people perform this function in the first place.

Respondents were asked whether they would prefer to receive a number of services from AI acting alone, from a human supported by AI, or from a human acting alone.

AI beat humans for two topics: The highest AI margin was for “provide show or movie recommendations,” where 35% preferred AI alone, compared to 27% who wanted human-only recommendations. (38% preferred AI-augmented human advice.) The second-highest AI margin was for “improve knowledge and/or skills around a specific topic,” where 30% preferred AI alone, 25% preferred humans alone, and 46% preferred AI-augmented humans.

For elementary school education, my own current preference is for the curriculum to be taught in a fully individualized manner by AI, but with human teachers in the classroom to serve as role models and keep order. For business professionals, on the other hand, AI is getting to the point where it’s probably the best way to learn many job skills.

The worst AI score was for “help diagnose health/medical problems,” where only 12% preferred AI alone. In contrast, 43% preferred humans alone, and 45% preferred AI-augmented humans. (This despite much research showing that AI often surpasses human doctors at medical diagnosis.)

For now, users are mostly interested in getting help from AI alone without any human involvement for lower-stakes decisions, such as what show to watch. For higher-stakes decisions, they still want a human involved but are mostly open to AI augmentation for that human. (NotebookLM)

The deeper pattern is not merely that AI wins low-stakes tasks and humans win high-stakes tasks. People are more willing to accept AI alone when the decision is reversible, and nobody needs to be clearly accountable for the outcome. They want humans in the loop when the costs of error are consequential, emotional, or socially exposed. That is a useful design rule for AI products: the higher the stakes, the more the interface must support oversight, explanation, and recourse.

Another low-AI category was outfit selection: for “help you pick out an outfit,” 22% preferred AI alone, 45% preferred humans alone, and 33% preferred AI-augmented humans. My guess is that people treat clothing as a social and expressive judgment, not merely an optimization problem. Even as AI gets better at visual taste, many users will still want a human mirror in the loop.

Outfit advice from AI? Not yet. (NotebookLM)

This highlights a crucial UX threshold: the divide between utility and identity. Users are increasingly eager to outsource utilitarian, low-stakes tasks (like data analysis or movie recommendations) to AI, but they fiercely protect domains tied to self-expression, personal taste, and social signaling. Choosing an outfit isn’t just a computational problem; it’s an act of human identity. Designing AI for these identity-driven tasks requires shifting the UX paradigm from “efficient automation” to “collaborative inspiration.”

And that may be the real theme of this week’s roundup. AI is moving out of the novelty phase and into the workflow phase. As adoption spreads, the winners will be judged less by raw model capability and more by whether they preserve context, reduce interaction cost, and earn enough trust to be used for consequential work. In the long run, usability is not an afterthought to AI capability. It is the differentiator.

About the Author

Jakob Nielsen, Ph.D., is a usability pioneer with 43 years experience in UX and the Founder of UX Tigers. He founded the discount usability movement for fast and cheap iterative design, including heuristic evaluation and the 10 usability heuristics. He formulated the eponymous Jakob’s Law of the Internet User Experience. Named “the king of usability” by Internet Magazine, “the guru of Web page usability” by The New York Times, and “the next best thing to a true time machine” by USA Today.

Previously, Dr. Nielsen was a Sun Microsystems Distinguished Engineer and a Member of Research Staff at Bell Communications Research, the branch of Bell Labs owned by the Regional Bell Operating Companies. He is the author of 8 books, including the best-selling Designing Web Usability: The Practice of Simplicity (published in 22 languages), the foundational Usability Engineering (30,030 citations in Google Scholar), and the pioneering Hypertext and Hypermedia (published two years before the Web launched).

Dr. Nielsen holds 79 United States patents, mainly on making the Internet easier to use. He received the Lifetime Achievement Award for Human–Computer Interaction Practice from ACM SIGCHI and was named a “Titan of Human Factors” by the Human Factors and Ergonomics Society.

· Subscribe to Jakob’s newsletter to get the full text of new articles emailed to you as soon as they are published.

· Follow Jakob on LinkedIn.

· Read: article about Jakob Nielsen’s career in UX

· Watch: Jakob Nielsen’s first 41 years in UX (8 min. video)

Summary: AI can translate a single source across media, styles, and audience frames. That makes it possible to reuse core material in formats better suited to different platforms and audiences. Shakespeare’s Hamlet serves here as a case study in a wide range of AI media transformations.

Thinking about AI as a media form is to recognize that it’s not a single media form, but rather all of them. AI enables creators to transform any content from one media type into another. An article becomes a comic strip becomes a training video becomes a song. All in a few minutes’ work.

To showcase this AI-native content thinking, I decided on a case study of a single piece of content made into many other formats: William Shakespeare’s classic play Hamlet, Prince of Denmark.

This case study remixes Hamlet along three separate axes. Sometimes the medium changes while the story stays mostly intact, as in the comic or music video. Sometimes the style or setting changes, as in the Viking adaptation or Skaldic poem. Sometimes the intended audience changes, as in the tourist posters, tabloids, law review, or consultant decks. AI makes all three kinds of transformation cheap enough that the limiting factor becomes editorial judgment rather than production cost.

Hamlet is probably one of the top three works of world literature, together with Homer’s The Iliad and Lady Murasaki’s Tale of Genji. You likely know the story already, but for a refresher, I recommend reading my 21-page comic Hamlet: The Comic Book. Since this is a long comic, I published it as a separate web page, rather than including it in this article.

Hamlet hardly needs introduction, but if you want a refresher, I recommend my 21-page comic Hamlet: The Comic Book, which I published separately because of its length. (That article also includes a behind-the-scenes account of my workflow.)

At 21 pages, it is my most ambitious comic yet, surpassing my 18-page comic on the history of graphical user interfaces.

Hamlet is a particularly good stress test for AI media transformation. It has a plot many readers already know, sharply differentiated characters, memorable lines, strong visual scenes, and enough ambiguity to support tragedy, parody, analysis, and pastiche. In other words, it remains recognizable even when compressed, relocated, or stylistically bent.

Film

Maybe the most obvious media format for transforming a theatrical play is a film. However, I didn’t make one since there are already many good movie adaptations of Hamlet, including classics starring great actors like Sir Lawrence Olivier, whom we will not be able to surpass with AI for another two to three years.

You can watch the AI short video version I made of another Shakespeare play, Pericles, Prince of Tyre (YouTube, 3 min.), to see what’s possible with almost-current AI video models. I made the Pericles movie almost 8 months ago with Veo 3.0, and I could probably do somewhat better today with Seedance 2.0, which (a) offers 15-second clips instead of 8 seconds, and also (b) allows better direction of multi-shot sequences with an omni-reference model.

Viking Comic Book

The 21-page comic plays Hamlet straight: Renaissance Denmark, circa 1601, staged at Elsinore Castle. Since Elsinore is usually identified with Kronborg Castle in Helsingør, Denmark, I used Kronborg as the model for my sets.

But AI doesn’t require you to convert content literally, or according to tradition. We can morph it as we please.

To make this point, I also created a short 5-page comic strip set in Viking-era Denmark. Shakespeare copied his plot from the legend of the Viking Prince Amleth in the medieval Danish author Saxo Grammaticus’s book Gesta Danorum (the Vita Amlethi episodes), which include all the key elements, such as a prince’s father murdered by his brother, who then marries the widowed queen; prince feigns madness to survive; is sent to Britain with a death order he reverses; returns to kill the usurping uncle. As Steve Jobs said, great artists steal. Shakespeare’s story is more elaborate than Saxo’s, but I do find it appropriate to set this comic strip in Viking Denmark, to honor the original.

Every medium preserves some things and discards others. Comics are good at staging and facial expression. Poetry compresses motive into image and sound. Data visualization reveals structure while sacrificing emotional nuance. Posters and magazine covers thrive by reducing the material to one sharp angle. That is why media transformation is not just repackaging. It is a way of discovering what a source text can survive, and what each medium makes newly visible.

Poetry

Continuing on the Viking theme, let’s transform Shakespeare into Skaldic poetry:

The Drápa of the Dane-Prince

1. The Ghost’s Tale

Grim the ghost of gloaming
Greets the heir of Denmark.
“Snake-tooth slew the sleeper,
Stole the seat of rulers.
Brother-bane has broken
Bone-house of the king-lord.
Hate-full hero, hear me:
Hew the hand of murder!”

2. The Rat and the Trap

Madness masks the mind-wolf,
Mirthless game is plotted.
Rat in room of rushes
Reaps the reed-of-battle.
Caller of the conscience
Catching king in play-act;
Fly to foreign far-lands,
Fear the foe is waiting.

3. The Sword-Storm

Willow-wife of waters
Wanders to the drowning.
Cup of curse is carried,
Cold the queen has drunk it.
Venom vexed the victor,
Vengeance strikes the slayer.
Spear-storm ends the story,
Silence takes the princes.

I used Gemini 3 Pro Deep Think for this poem. Gemini aimed to mimic dróttkvætt, the court meter of Icelandic skalds such as Egill Skallagrímsson, adapted here for English. The key constraints are these:

1. Syllable Count: Every line has exactly six syllables (e.g., Grim the ghost of gloam-ing).

2. Trochaic Rhythm: The lines generally end with a trochee (a stressed syllable followed by an unstressed one), creating a falling rhythm distinct to the style (e.g., Den-mark, rul-ers, mur-der).

3. Alliteration: The lines are linked in pairs by “staves” (alliterating sounds):

   • Odd lines (1, 3, 5, 7) contain two alliterating sounds.

   • Even lines (2, 4, 6, 8) begin with a “head stave”—the first stressed syllable must match the sound of the previous line.

Key to Kennings

Skaldic poetry relies on kennings: compound metaphors that act as riddles to describe people or objects.

• Ghost of gloaming: The Spirit of King Hamlet (gloaming is twilight/darkness).

• Snake-tooth: Poison (referencing the “serpent” that stung the King).

• Seat of rulers: The Throne.

• Brother-bane: Claudius (the killer of his own brother).

• Bone-house: The Body (a classic Old Norse kenning for the physical frame).

• Mind-wolf: Hamlet (a warrior of the intellect; implies danger and ferocity).

• Rat in room of rushes: Polonius (Hamlet calls him a rat; he hides behind a tapestry in a room with rush-covered floors).

• Reed-of-battle: A Sword (a plant/stalk that grows in the field of war).

• Willow-wife: Ophelia (associated with the willow tree where she died).

• Spear-storm: Battle/Duel (a weather metaphor for the chaos of fighting).

Music Video

I made a music video based on Hamlet (YouTube, 4 min.) The song was made with Suno 5.5, and the animations with the Chinese video model Seedance 2. Both are excellent AI content creation models, and probably the best in the world for their respective media forms. I think this music video came out well, but give it a listen.

I spent US $160 on Seedance to animate this 4-minute video. While $40 per minute is steep for an amateur project, it’s a bargain compared to traditional indie music videos ($1,000–$5,000/min) or major label productions ($25,000+/min). I could have spent closer to $100, but because AI video generation is unpredictable and Seedance takes about 10 minutes per clip, I paid an “impatience tax.” Rather than reviewing clips linearly, I fired off 3–4 parallel generations per scene to ensure at least one aligned with my directorial vision. For instance, when prompting for the traveling troupe of theater actors arriving in wagons, three of the four generated videos inexplicably showed the wagons backing out of the castle. Running parallel variations costs extra, but bypassing a sequential workflow saves hours of trial-and-error.

The production budget for my Hamlet music video was higher than it needed to have been if I had been more patient waiting for Seedance to finish generating clips and only paid for additional clips when my first attempt failed.

One interesting new tradeoff appears in AI production: money often buys latency reduction rather than better craftsmanship. Parallel generation lets a solo creator compress waiting time, even when the final clip is no better than what a slower workflow might eventually have produced.

Data Visualization

Let’s first have AI count the number of lines spoken by each character in each act and then ask it to visualize this data:

It makes sense to have the protagonist’s share of the lines grow by each act as the plot thickens, but it was a bold choice on Shakespeare’s side to follow Hamlet’s dominance point in Act III by placing him mostly offstage in Act IV. The visualization clearly shows this dramatic shift.

Another data visualization by Nano Banana Pro was less successful:

The second visualization is less successful. Edward Tufte would call it chart junk, but ornament is not the main problem. The real issues are accuracy and clarity. It wrongly gives Ophelia lines in Act V, even though she dies in Act IV. It also breaks consistency by changing the color coding for Hamlet across acts, making the graphic harder to follow. And it misses the most interesting story in the data: the protagonist dominance index, Hamlet’s rising and then collapsing share of the dialogue.

This visualization example is a reminder that AI can generate plausible analytical graphics before it can generate trustworthy ones. The human job is no longer to draw the chart by hand; it is to verify the counts, inspect the encoding, and reject attractive nonsense. AI lowers production cost, but it raises the premium on editorial judgment and quality control.

Slide Decks

How would Hamlet’s famous “To be, or not to be?” soliloquy work as a PowerPoint presentation by a consultant? Let us find out: first, a business strategy consultant, and second, a UX consultant take on this assignment. After all, job requirement number one for consultants is to be able to bamboozle client management with fancy slides, even when you don’t have much substance to go on.

What makes this transformation fascinating is how the medium inherently alters the psychological framing of the content. Hamlet’s soliloquy is a raw exploration of existential dread and paralysis. However, when forced into the rigid format of a consulting slide deck, that profound human emotion is automatically sanitized into sterile, actionable corporate jargon like “Go/No-Go Heuristic Frameworks” and “Strategic Pivots.” This demonstrates that AI isn’t just translating words; it is actively adopting the cultural mindset of the target medium. (This is to your advantage when you want that alternative approach.)

Posters

The next transformations shift from retelling the story to targeting different audiences. A tourist poster treats Hamlet as place branding. A tabloid cover treats it as scandal. A law review turns it into a problem of legitimacy, inheritance, and culpability. A trade magazine filters it through the concerns of a profession. AI is not just changing format here. It is changing the editorial angle.

Posters are a media format optimized for punchy, singular messages. Here are two tourist posters I made for visiting Denmark and seeing Hamlet’s castle.

Tourist poster, 1960s Modern Scandinavian style.

Modern tourist poster.

Newspapers and Magazines

Newspaper front pages and magazine covers are similarly high-impact formats, but they allow more detail than posters and make the editorial angle explicit. Each genre reframes the same story for a different audience.

Clockwise from upper left: Women’s lifestyle, men’s lifestyle, bodybuilding, teen lifestyle.

Clockwise from upper left: Celebrity rag, mainstream news magazine, royalty gossip, supermarket tabloid.

Clockwise from upper left: Law review, fashion magazine, UX professionals’ trade magazine, physicians’ trade magazine.

Clockwise from upper left: Gardening, Popular Mechanics, business magazine, sports.

Newspaper front page.

Funny Cartoons

If you subscribe to my newsletter, you know that I am a great fan of funny comics. Here are three based on Hamlet:

Miscellaneous Media Formats

Here’s Hamlet as a snow globe, as a segment from the medieval Bayeux Tapestry (the original of which records the invasion of England by William the Conqueror in 1066), as a science fiction paperback, a collectible doll, and a social media feed.

Hamlet in a snow globe.

Abbreviated plot of Hamlet as a segment of the medieval Bayeux Tapestry. In keeping with the actual tapestry, the Banana wrote the captions in Latin. (Translated, they mean: [1] “Here Hamlet sees the ghost of his father.” [2] “Play of mice,” meaning the Mousetrap play-within-the-play. [3] “Hamlet kills Polonius.” [4] “Ophelia dies in the river.” [5]: “Final duel, and the death of everyone.”) Note the use of the letter “V” for “U,” which is consistent with Latin typography.

This second science fiction paperback book cover was made with the GPT Image 1.5 image model. All the remaining images in this article were made with Nano Banana (Pro or 2 versions).

Ophelia doll.

Prince Hamlet’s social media posts on X (formerly Twitter).

Non-Fiction Storytelling

Finally, AI can repurpose the logic of a fictional work into non-fiction. Here I used Hamlet to tell a product-development story about the risk of trusting analytics over user observation.

I made two versions of the comic: one set literally in the modern workplace, the other staged in Renaissance costume inside a present-day company. While absurd, the Renaissance version works better because it preserves immediate recognition of the Shakespearean roles while still carrying the modern analogy.

That is the larger implication for content strategy. One source asset can now generate explanatory, promotional, analytical, and entertaining derivatives at a scale that once required separate teams. But the winning strategy is not to spray the same material across channels. It is to ask what each medium is uniquely good at, then rebuild the content around that strength. AI removes much of the production friction; it does not remove the need for taste, selection, and verification.