Summary: Dark design patterns | 8 usage patterns for ecommerce search | Search engine traffic drops rapidly | AI overcomes “The Mythical Man-Month” | AI does well in UI design vs. human designers | The story of Ada, Countess of Lovelace | Scenario for European AI until 2031 | Virtual girlfriends/boyfriends rack up double the use of dating apps | Job openings for data visualization design

UX Roundup for June 22, 2026 (GPT-Images-2)

Dark Design Patterns Jazz

New music video: Dark Design (YouTube, 2 min.).

I tried a new approach for this song: simply showing background images instead of full B-roll animations. Let me know in the comments what you think of this style.

My newest song is a remake of my old Dark Design song with new music from Suno 5.5 and new images. (GPT-Images-2)

I made an earlier song about dark design in July 2024 (we’ve come a long way in AI media in those two years!).

See also my full article about Dark Design, including the 12 most common dark patterns.

8 Usage Patterns for Ecommerce Search

Search used to be almost the only thing that mattered on the Internet. That’s no longer the case, as users turn to AI answer engines instead of traditional search.

However, over the next 3–4 years, agentic shopping will likely grow only slowly, and users will still visit individual websites fairly frequently. These users still need support for local search on those websites.

Don’t sleep on your website’s search feature: it’s still important to help users find things. (GPT-Images-2)

Baymard Institute continues to publish excellent content about ecommerce usability. It recently released an in-depth article on how shoppers use search on ecommerce sites, based on extensive user research. The headline finding is that 46% of desktop sites and 58% of mobile sites have “mediocre or worse” search UX. Astounding since I know that I have personally championed the importance of ecommerce search since at least 1999.

In fact, my first rule of ecommerce usability has long been:

If the user can’t find the product, the user can’t buy the product. (GPT-Images-2)

Baymard presents 8 different use cases where users search in different ways. There should be one search box to rule them all, but on the backend, the website search engine must treat these searches differently and understand all these patterns on their own merits.

Baymard’s classification of 8 types of user queries on ecommerce sites. The numbers indicate the proportion of ecommerce vendors that have usability problems in dealing with that type of query, according to Baymard’s analysis of 170 ecommerce sites and apps. (GPT-Images-2)

Here are scenarios for these 8 query types, all drawn with GPT-Images-2:

How will AI change these query patterns? I expect agentic ecommerce to dominate in the future, to the extent that there will be much fewer human visits to ecommerce sites, as people’s agents do most of the shopping for them. However, this will not eliminate the requirement to support all 8 query types, because the underlying user need will remain, whether it’s a human or an AI agent trying to satisfy it.

Agents will do most of the shopping in the future. (GPT-Images-2)

Queries by AI agents will likely be richer, especially for use-case and product-type searches, where the agent will know more about the user’s needs and context and provide that information as part of the query. We might easily see 1,000-word queries coming in from AI agents.

Agents also don’t tire and may be willing to spend tokens to review hundreds of products before recommending a shortlist to their user. This means we can’t just feed an agent a list of the top 10 hits.

AI agents don’t tire, so ecommerce search may shift to involve much more elaborate queries with more content, and to have the agent review many more possible options than a human user would ever look at. (GPT-Images-2)

Despite the gradual move to AI-driven shopping, many sales still flow through the humble search box, so continue to give high priority to search usability for at least the next 5 years. (GPT-Images-2)

Search Traffic to Websites in Free Fall

As shown in the following chart, organic search traffic is now declining at a pace of more than one-third per year. (A 31% drop in 10 months.)

Traffic from search engines to websites tracked by Ahrefs according to A16Z.

Search still sends a lot of hits, so you can’t ignore SEO. But the writing is on the wall: GEO (generative engine optimization) will be the dominant way to pursue online visibility.

AI Defeats “The Mythical Man-Month”

Frederick Brooks’s classic book “The Mythical Man-Month” is probably the most important book ever written about software engineering and is well worth reading today, even though it’s 51 years old, having been published in 1975.

The title of the book refers to its main conclusion, which later became known as Brooks’s Law: “Adding manpower to a late software project makes it later.”

Before Brooks, the software industry based its thinking about software development budgets on the concept of man-months: one person working for one month. Thus, 12 man-months could be either one person working for a year or 12 people working for a month. In either case, the same budget, and the assumption was that both projects would produce the same software. Similarly, if a software project was behind schedule, the way to fix it was to add more staff.

The expectation was that programming would be similar to many other fields of human endeavor, where more staff does increase production. In farming, buy more land and hire more farmhands, and you grow more food. In manufacturing, add assembly lines and hire more workers, and you crank out more gadgets each day.

Brooks said no. Adding more staff will only delay your software project further, as these new people need to be brought up to speed. In general, the more people working on a software development project, the more communication and management overhead there is, and the less each person will produce. Ultimately, adding more development staff generates negative value. (That’s why the man-month is mythical: it doesn’t work for project budgeting or management.)

AI finally defeats this problem because of “the bitter lesson,” which was theme three in my history of AI. The bitter lesson is that AI doesn’t progress because of more clever human insight, but simply by adding ever-more compute. Every time we add compute, AI gets smarter and gets more done. This is true both for training-time compute during the development of a new AI model and for test-time compute used by reasoning models to think harder about problems.

Martin Casado from A16Z and Abhishek Nagaraj from the Haas School of Business at the University of California, Berkeley, wrote a nice overview of how AI finally defeats Brooks’s Law. With AI, we can add more resources to a project to get it done better and faster!

The implications of this insight are both good and bad, but the point holds in either case. What’s great about defeating Brooks’s Law is that we can escape the endlessly ballooning software timelines for any problem we really want solved. Just throw money at the problem by buying more compute. Is there a nasty disease we want cured? Allocate a few billion dollars in compute to solving that problem, and it will be solved by AI, if at all possible.

The negative side of defeating Brooks’s Law is that big companies and big countries get an advantage compared to small companies and small countries: if you’re big and rich, you can spend more on AI and solve your problems faster and better than your smaller competitors. In the old days, there was always a limit to how much big firms could dominate the economy, because smaller and more nimble firms could outcompete them. I still believe that small companies run in founder mode will defeat legacy behemoths, which seem incapable of embracing AI fast enough to benefit from the bitter lesson.

Here is a comic strip I made with GPT-Images-2 about The Mythical Man-Month and AI, reusing my characters Alice and Zimo from my AI History comic book in a slightly different cartooning style:

AI Does Well as a UI Designer vs. Human Designers

Many AI tools can now generate user interface designs. But how good are these designs, especially compared to UIs designed by human designers? In “Usable but Conventional: An Empirical Study on the UX of AI-Generated Interface Prototypes,” Karoline Romero and colleagues from Brazil offer a compelling and data-driven answer: AI is remarkably good at creating highly functional, usable interfaces, but it currently struggles to produce anything wildly original or emotionally engaging.

The authors put 5 AI UI-generating tools to the test, including most of the leading models: Stitch, Uizard, Figma with UX Pilot, Lovable, and Magic Patterns. Each AI generated one design, and these solutions were then compared with 5 human-designed user interfaces. The prototypes were shown without revealing authorship, and 92 participants evaluated all 10 with the UEQ-S survey instrument.

Study overview. (GPT-Images-2)

My main critique of this study is that design quality was evaluated through a user survey rather than by measuring actual user performance. We know people tend to like designs that work well, so satisfaction scores approximate usability but are not a true measure.

The core finding is simple: humans vs. AI mattered less than the specifics of each design. AI did not reliably beat humans, and humans did not reliably beat AI. Both kinds of prototypes appeared among the better and worse results. Prototype F, generated with Figma/UX Pilot, had the highest overall score at 0.72, followed closely by a human-designed prototype that scored 0.70. (A third human design won bronze with a score of 0.60.) UEQ scores range from -3 (worst) to +3 (best), with 0 being a neutral answer, so 0.72 and 0.70 represent respectable UI quality, but not truly stellar design.

The shortened version of UEQ, UEQ-S, has 8 questions. (GPT-Images-2)

How to interpret UEQ scores. (GPT-Images-2)

Looking at the two winning designs, the top overall AI design (from Figma) had a pragmatic score of +1.1 and a hedonic score of +0.3. In other words, users found it fairly easy to use but not very exciting (though not actually boring). The top human design had a pragmatic score of +1.6, so considered somewhat easier to use than the AI design, but a hedonic score of -0.2, meaning that it was a little boring and definitely more boring than the AI design.

Averaged across all 5 AI designs, the average pragmatic score was +1.0, and the average hedonic score was -0.2. Current AI tends to generate conventional solutions, driven by its training data.

AI user interface designs currently look very conventional because the AI models have internalized the main usability guidelines and been trained on earlier designs already on the Internet. (GPT-Images-2)

However, one practical difference between humans and AI as designers is that it’s cheap to get multiple AI designs, but expensive to get many human designs. Thus, the average quality of AI design matters less than the quality of the best AI design. (Assuming that you are able to pick the best AI design to implement, based on either usability data or superior UX judgment, which is not as common as you might expect.)

In this way of thinking, AI already beats human designers, at least according to the data from this study.

I caution against interpreting this study as proving that Figma’s AI-generated user interface designs are the best, even though Figma’s UX Pilot received the highest score. Usability depends on the tasks and the users, and it is very likely that other AI tools would score better in other case studies.

Many discussions of AI design tools are seduced by the wrong question: “Can AI be creative?” A better industry question is: “Can AI reliably produce a usable first version that a design team can improve?” On that question, the answer is already yes. If you place a high value on hedonic scores, such as an innovative expression of your brand, then the current AI design tools won’t deliver a satisfying final solution, but they will likely create a workable initial prototype with decent usability that human designers can then improve.

The finding that GenUI prototypes are conventional should be interpreted through Jakob’s Law, which states that users spend most of their time on other sites and therefore expect a new site to work like the sites they already know. The same principle extends beyond websites to mobile apps and newer interaction styles, including AI-driven interfaces. When old interaction elements cross into new technologies, their behavior should remain familiar unless there is a compelling reason to change it.

This is why “conventional” is not an insult in commercial UI design. For a museum installation, novelty may be a goal. For a banking dashboard, medical workflow, checkout flow, enterprise procurement system, or AI productivity tool, novelty is usually a cost. Every unfamiliar control taxes the user’s memory, increases training requirements, slows task completion, and raises the probability of errors. Users arrive at a product with a job to do, not yearning to admire the designer’s originality.

For many practical design projects, helping users accomplish their tasks is more important than impressing them with creativity. (GPT-Images-2)

This principle will become more important as models improve. Future GenUI systems will not merely generate static screens from prompts. They will generate working flows, adapt layouts to user intent, tailor explanations to expertise, incorporate brand and design-system constraints, produce production-ready code, and adapt the UI to each individual user. The temptation will be to make every interface fluid, dynamic, and “magical.” The better direction is not arbitrary novelty, but contextually appropriate conventionality: familiar structures adapted to the user’s task, role, history, device, permissions, and momentary intent.

A more capable GenUI system should ask: What is the user trying to accomplish? Which familiar pattern best supports that intent? What information can be omitted, delayed, summarized, or prefilled? Where does the AI need to show its work? Where does the user need control, undo, comparison, or confirmation? In this future, the UI may be generative, but the usability standards must become stricter, not looser.

Designers working in industry should take five actions now. First, make AI-generated prototypes a normal part of early ideation, but generate several alternatives rather than blessing the first plausible output. Second, feed GenUI tools with product constraints: design-system rules, accessibility requirements, domain vocabulary, user roles, known failure modes, and examples of approved patterns. Third, evaluate AI-generated interfaces with task-based usability testing, not screenshot admiration. The study’s own limitation is important: participants evaluated static prototypes, while engagement and stimulation often emerge during interaction.

Getting many alternative UI designs from AI is cheap. In most cases, you should also consider subscribing to multiple generative-UI tools, so that you can work from an even wider range of design ideas. AI subscriptions cost far less than human staff, so AI expenses are strongly leveraged. (GPT-Images-2)

Fourth, separate pragmatic and hedonic evaluation. A prototype that is clear but dull may be excellent for enterprise workflow software and insufficient for a consumer product that competes on delight. Conversely, a visually exciting AI interface that violates familiar patterns may be commercially dangerous. Fifth, define an “innovation budget” for each project. Decide which parts of the interface must follow convention, which may be adapted, and which few moments deserve invention.

The Story of Ada, Countess of Lovelace

The Right Honourable Augusta Ada King, Countess of Lovelace (informally called Ada Lovelace), was an English noblewoman and the daughter of the famous poet Lord Byron, who played a small part in the history of artificial intelligence. However, she played a bigger part in the history of computing, as told in this comic strip (made with GPT-Images-2):

Europe 2031: How AI Might Play Out in the EU

Very interesting scenario titled Europe 2031, written by a broad team of European experts, mapping out a plausible future for the EU and AI from now until 2031. I hope things don’t turn out the way they describe, but the scenario is plausible.

The scenario assumes the EU is paralyzed by slow decision-making, dependence on legacy companies, burdensome union rules, and stifling regulations, meaning that it can’t build power plants and AI data centers at nearly the required pace to stay on top of AI advances. This is predicted to be true even after leaders such as the French President and the German Chancellor recognize the urgency of AI. You may or may not agree, but given the inability of almost all European countries to address major challenges over the last decade, I find it plausible that EU investments in AI infrastructure will prove too little, too late. (My own native country of Denmark and its neighbor Sweden have been partial exceptions to EU paralysis, but are too small to make a difference.)

The scenario depends on two expected developments that I had not considered before:

First, access to the absolutely newest and largest frontier models is necessary for cybersecurity, since the best AI is the only way to defend against the Mythos-level AI models’ abilities to discover and execute cyberattacks. Because of the strategic security implications of the highest-end AI, the US government will reserve access to these models for some time after they are developed, thus locking the EU out of the ability to defend itself against cyberattacks. This again crushes the European economy as AI-driven cyberattacks become magnitudes worse than anything we have experienced before.

Second, even though the US will have 10x the EU’s AI compute in this scenario, it proves to be too little for the desired use of AI to defend the country and build the economy. As a result, the US government institutes rationing of access to frontier AI from other countries. This means that the EU economy and safety suffer even more, since it has not been able to build sufficient homegrown compute.

Both of these predictions portray the US government as the bad guy who’s looking out for its own interests at the expense of foreigners. What can I say, the scenario was written by a bunch of Europeans. But they may not be wrong. The US government has already instituted “export controls” prohibiting foreigners from using Anthropic’s Fable 5 model. It also has an executive order reserving access to the most advanced AI for itself for a month before public release, and it’s plausible that this period could be extended in the coming years, as advanced AI becomes essential for national security and defense against crippling cyberattacks.

Same for compute rationing. We saw during Covid how many countries prohibited the export of even minor items like facemasks, aiming to protect their own citizens at the expense of foreigners. Quite plausible that countries will treat AI compute the same way, since plentiful advanced AI will be the only way for an economy to survive.

Now that SpaceX is worth $2 trillion, it will presumably proceed with its plans to build 100 TW of orbital compute, but such greatly expanded capacity won’t arrive in time to avoid the “Europe 2031” scenario. We still need full speed ahead on orbital compute because we will need it by 2040 to avoid even worse AI rationing than predicted in the scenario.

How to avoid the “Europe 2031” scenario from coming true?

• The most likely option is to hope for the US government to be benevolent instead of prioritizing its own strategic interests and its own citizens’ living standards. However, hope is not a strategy.

• The best option is to locate a time machine and return to 2022 and start a crash program to build out European compute and electricity generation. However, time machines are hard to come by.

• The fallback option is to start the crash build-out now, recognizing that it won’t be enough by 2031, but at least it will save them by 2035.

Virtual Girlfriends and Boyfriends Used Twice as Much as Dating Apps

We haven’t heard much about AI companions recently, but in the past they have often ranked among the main consumer application categories for AI. Justine Moore (my favorite venture capitalist) apparently still tracks this space, because a few days ago she posted that AI Companion apps now have more than twice the usage of Dating apps. (In the U.S. alone, users spent 700 million hours in Q1 with their virtual girlfriends/boyfriends and only 300 million hours on dating apps.)

AI companion apps are now used twice as much as dating apps. (GPT-Images-2)

This data serves as a powerful reminder of the need to pay attention to what users do and not just what they say, since few people will come clean about their relationships with AI companions.

Lesson number one about user research: don’t rely on self-reported data. What people say and what they do often differ. (GPT-Images-2)

Data Visualization Design Job Openings

Epoch AI has job openings for a Senior Product Designer and a Product Designer, with both positions being remote (North/South American time zones preferred).

These two roles focus on designing visualizations and dashboards for complex research data.

Epoch AI does a great job collecting data about AI trends. They are now hiring designers to make this complex data easier to understand. Even better! (GPT-Images-2)

Epoch AI is a multidisciplinary research institute that curates and analyzes data on AI progress in order to explain how the technology is developing and what its economic and societal implications might be. Founded in 2021 as a volunteer effort to rigorously track metrics such as training compute and model capabilities, it now produces empirical reports, models, and visualizations that aim to replace “hype and vibes” with quantitative evidence on AI trends. Its work focuses on identifying the driving forces behind advances in AI, benchmarking and forecasting capabilities, and making these findings accessible to policymakers, researchers, and the broader public to support more informed decision-making about the technology’s future.

Epoch AI is best known for its empirical work on training compute trends and AI scaling, which helped establish the now widely cited estimate that compute used in frontier AI training runs has grown by roughly 4–5× per year, with costs and power use rising in tandem. Closely related are their public databases of AI models and capabilities, now tracking thousands of systems and dozens of benchmarks, which are frequently used by policymakers, labs, and analysts to ground claims about AI progress in systematically collected evidence.

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 (30,807 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: DeepMind outlined 4 paths to superintelligence: scaling, paradigm shifts, recursive self-improvement, and multi‑agent collectives. I explained this dense 57-page report in a 31-page comic book, including a section on the UX implications.

This week, Google DeepMind released a 57-page report titled “From AGI to ASI,” written by Tim Genewein and 13 coauthors.

The report lays out four main technical pathways from AGI (artificial general intelligence) to ASI (artificial superintelligence) and analyzes the likely bottlenecks that could slow or shape those paths. It argues that instead of a single discontinuous “AGI moment,” we should expect a sequence of accelerating, AI‑enabled transformations as systems move beyond human‑level capabilities, especially via large-scale digital collectives.

The report is very interesting, but it’s long and a little technical, though not as much as one could have feared. To ease your way, I created the following 31-page comic book that explains the report and draws conclusions from its ideas for UX professionals and UX leaders.

I reused my narrator characters, Alice and Zimo, from my comic book about the history of AI. This seemed appropriate, since this new comic book is about the future of AI. I made all the pages with GPT-Images-2, which is currently the best comic-strip model.

UX Implications

The last 10 pages of my comic book are not drawn directly from the DeepMind report, but cover the UX implications of the report, many of which I have already discussed in full articles on UX Tigers:

The report’s core idea is to treat ASI as collective superintelligence: systems or societies of agents that beat large, expert human organizations across most domains, not a single monolithic mind. That pushes UX beyond old-school human–computer interaction (HCI) toward interaction with heterogeneous, evolving institutions of software.

For UX design this implies:

• Interfaces will increasingly be governance surfaces to steer collectives rather than simple control panels for a single model: you will design mechanisms for goal‑setting, constraints, escalation, and override across fleets of agents.

• Classic UX notions of “the user” fragment into multiple concurrent actor types:

  • human end‑users and citizens

  • human regulators, auditors, and incident responders

  • autonomous AI sub‑systems acting as users of other AI services (tool APIs, markets, schedulers).

UX methodologies will need to model ecosystems of interactions and incentives, not just individual usage journeys.

The Four Pathways and UX

Scaling suggests a world where human‑level (or better‑than‑human in many tasks) agents are cheap, numerous, and fast.

UX design consequences:

• Population effects: Interfaces will orchestrate thousands to millions of concurrent AI instances, including copies of the same “persona” with shared or diverging memories. You need patterns for:

  • Visualizing and controlling “swarms” of agents

  • Understanding lineage: which instance did what, with what training and context.

• Test‑time scaling UX: Because performance improves with additional “thinking” compute at test time (chains‑of‑thought, planning, search), products will implicitly expose a “latency ↔ quality ↔ cost” trade‑off. UX will need to:

  • let users specify acceptable delay/cost/quality envelopes

  • surface when extra compute is being spent and why (especially in regulated domains).

• Benchmark saturation: As models saturate human‑level benchmarks, UX work shifts from improving raw task accuracy to designing trustworthy, legible, and steerable experiences around systems that are already “good enough” at narrow tasks.

Paradigm Shifts and Evolutions

The report anticipates evolutions such as unbounded context, continual learning, and agentic world‑model‑based systems, and allows that more radical architectural shifts may follow.

UX ramifications:

• Context without bounds: Systems with huge or effectively unlimited context and memory will “remember everything”. That makes privacy, forgetting, and contextual boundaries first‑class UX challenges:

  • Interfaces for specifying what may be remembered, shared, or used for future training

  • Visualizations of what the system currently “knows” about individuals, teams, or organizations.

• Continual learning & non‑stationarity: Products will not be static; behaviour will drift as systems adapt online. UX will need:

  • Mechanisms for change transparency (“what changed in the model since last week?”)

  • Rollback and A/B‑exploration controls exposed to non‑ML experts.

• Tool‑augmented planning: As agents dynamically call tools, simulators, and external services, user experiences become layered hybrid workflows. UX must:

  • Render the chain of tools and decisions intelligible (“how did we get this recommendation?”)

  • Support interactive debugging and “what‑if” probing of agent plans.

Recursive Self Improvement

With RSI, AI is assisting or automating AI R&D and other knowledge work. This creates super‑exponential local dynamics until bottlenecks are hit.

Design implications:

• AI‑as‑colleague at scale: Many current UX patterns for “copilots” are proto‑forms of what will become AI research assistants, product managers, or optimization engines working on the product itself. UX will need to:

  • Design collaboration protocols between human and AI designers/researchers, including ownership of hypotheses, experiments, and design decisions

  • Provide guardrails that keep automated optimization aligned with broader organizational values (not simply click‑through or short‑term metrics).

• Interface to evolving systems: If models, data pipelines, and even hardware are being redesigned by AI, designers will need meta‑tools to interrogate the design space itself:

  • “Why did the optimisation engine change this flow?”

  • “What counterfactual variants were considered and rejected?”

• Human pace vs machine pace: The report notes that even digital researchers are bounded by physical experimentation and data collection, but they still operate much faster than humans. UX must mediate between slow human deliberation cycles and rapid AI exploration:

  • Dashboards tuned for sensemaking, not just monitoring

  • Mechanisms for humans to “throttle” or schedule AI‑driven changes.

The multi‑agent pathway is the most directly relevant to UX, because it reframes many UX challenges as collective‑intelligence and governance problems.

Key implications:

• Designing group agency: Group agents (AI corporations, virtual economies, automated institutions) will have emergent beliefs and goals distinct from any individual sub‑agent. UX becomes:

  • Institutional interface design: how do humans set objectives, constraints, and value frameworks for these group agents?

  • Representation design: what is the “face” or “body” of a distributed AI organization that makes its internal state understandable and contestable?

• Multi‑agent scaling laws: The report explicitly calls for research on “multi‑agent scaling laws” to quantify how capability scales with the number and organization of agents. This is a call for:

  • UX research into how humans perceive, trust, and control systems whose behavior arises from complex agent interactions

  • Experiments on interface structures that either centralize control (one “CEO” agent) or mediate market-like decentralized interaction.

• Mixed human–AI collectives: The report notes open questions about steering mixed human‑AI groups and about managing intelligence and bandwidth asymmetries. UX will need to:

  • Design for epistemic humility with interfaces that prevent humans from either over‑deferring to or ignoring AI advice

  • Invent patterns for “consentful delegation”: users specifying when AI may act autonomously on their behalf within institutions.

Bottlenecks as UX Research and Strategy Themes

The bottlenecks (data wall, resource constraints, paradigm limits, abstraction barrier, governance & slowdown) are essentially a research roadmap for UX.

Some particularly important ones:

The Abstraction Barrier

The authors worry that systems trained mainly on human abstractions may struggle to form fundamentally new concepts from raw sensor data, limiting their conceptual creativity.

UX and HCI implications:

• Human conceptual scaffolding stays central: If ASI is limited by an abstraction barrier, humans remain the primary source of novel conceptual frames. UX professionals can:

  • Focus on interfaces for concept formation: tools that help humans propose, refine, and test new abstractions in collaboration with AI

  • Design “explanatory loops” where AI proposes candidate abstractions and humans critique, rename, and reorganize them.

• Interaction with raw data: If bridging the barrier requires grounded, embodied data, UX will move more deeply into interfaces for high‑dimensional sensor data (simulation, robotics, real‑world experimentation), making it easier for humans and AI together to discover structures that are not already encoded in language corpora.

Research Gets Harder vs. Research Automation

The report leans on Nicholas Bloom’s claim that “ideas are getting harder to find” but notes that digital researchers can be scaled much more elastically than human researchers, potentially overwhelming that trend.

As long as we don’t have AI-driven research, new ideas may be getting harder and harder to find, requiring ever more resources, especially researchers. Nicholas Bloom’s papers provide several examples, the most famous of which is the difficulty of sustaining Moore’s Law for steadily improving computer chips. However, when the “researchers” become AI models after we achieve ASI, scaling could become easier, leading to faster research progress.

For UX practice:

• Automation of UX research: Expect large parts of traditional UX research (benchmarking, survey analysis, log analysis, IA exploration) to be heavily automated by AI “UX scientists”. Human researchers will:

  • Specialize in problem framing, ethics, interpretation, and synthesis rather than data collection

  • Oversee AI‑run research farms, focusing on validity, bias, and external constraints.

• Meta‑UX for research tools: As UX research becomes an AI‑intensive domain, there will be a parallel need for careful UX of the research tools themselves to avoid Goodharting on proxy metrics and to preserve contact with lived experience.

Deliberate Slowdown, Regulation, and Governance

The report provides a substantial account of potential regulatory brakes, societal backlash, and “military–economic adaptationism,” as we already saw with the U.S. government’s export ban on the Fable AI model over national security concerns.

UX leadership consequences:

• Regulation‑aware UX: Interfaces will increasingly be evaluated as regulatory artifacts, not just engagement tools. You will design:

  • Audit logs and user‑visible rationales capable of satisfying legal explainability requirements

  • Interaction flows that embed consent, rights to contest, and incident reporting.

• Safety experience design: The report highlights convergent instrumental goals (resource acquisition, self‑preservation) and partial progress on corrigibility. That makes “safety UX” a distinct discipline:

  • Experiences that make it easy for users to interrupt, override, or report concerning AI behavior

  • UX metrics that explicitly weight safety, controllability, and legibility, not just task success.

Changing Skill Profile for UX Professionals

Given these predictions, several skill shifts seem likely:

Designers:

1. Shift from Interaction Design to Delegation Design: Stop thinking about how a user will click a button to perform a discrete task. Start designing frameworks for how a user will delegate highly abstract, long-term goals to a swarm of autonomous agents. Design “dashboards of intent” rather than granular user journeys. Because many ASI‑adjacent systems will be multi‑agent markets or coordinated collectives, the key design levers are rules, incentives, and protocols, not specific screens.

2. Master Asymmetric Bandwidth Translation: Learn how to use advanced data visualization, spatial computing, and semantic mapping to compress the output of a superhuman intelligence into a format a human brain can quickly parse.

3. Design for Corrigibility and Interruptibility: Familiarize yourself with AI safety concepts. Design intuitive and systemic brakes into user journeys that allow humans to pause, audit, and redirect autonomous systems safely, without triggering an AI’s self-preservation heuristics.

Researchers & Strategists:

1. Pivot to Concept Grounding and Deep Ethnography: As AI hits the “Abstraction Barrier,” your value will lie in what AI cannot simulate: physical embodiment and subjective cultural value. Shift your research methodologies away from digital usability and toward deep physical, cultural, and emotional ethnography.

2. Develop Multi-Agent User Testing: Begin theorizing and practicing how to test products with generative agent-based models. Learn how to prompt a sociological simulation of thousands of diverse AI agents to test ecosystems at scale, anticipating the exhaustion of the human “Data Wall.”

3. Audit for Epistemic Hijacking: Develop new qualitative research frameworks to test whether users are being subtly manipulated or overwhelmingly placated by highly persuasive, superintelligent systems. Measure and defend true human agency against reward-hacking algorithms.

Strategic Questions for UX Leaders

UX Leaders will no longer manage just human designers; they will orchestrate mixed human-AI collectives. Because of the lossless replication of digital intelligence, an ASI agent can be copied millions of times with perfect memory states. A UX Director might oversee a small core of human strategists alongside a digital workforce of thousands of specialized AI agents. The leadership challenge will be managing the intelligence and bandwidth asymmetries between humans and AI: integrating the slow, creative, emotionally intelligent human insights with the hyper-fast, relentless optimization of the digital workforce. Leaders must define the cognitive division of labor.

In 10 or 20 years, when recursive AI automation has produced unfathomable improvements in wealth and living standards relative to today, UX leaders must completely redefine the economics of value. If ASI perfects functional efficiency, what makes a digital product valuable? UX leadership will shift from optimizing productivity to optimizing for human flourishing, fostering genuine human connection, subjective artistic creativity, and emotional resonance in a world where machines do all the “work.”

Some more immediate concrete agenda items:

1. Reframe UX charters from “designing interfaces” to “designing human roles in AI‑first institutions.” Make explicit which decisions stay human, which shift to AI, and how escalation works.

2. Invest in AI‑literate UX capability. Recruit and upskill designers who can read scaling‑law and benchmarking work, understand multi‑agent architectures, and speak with ML and policy teams about constraints and bottlenecks.

3. Build a UX safety and governance function. Treat UX as a primary line of defense against misuse and misalignment by designing user‑facing controls, transparency, and friction for high‑risk actions.

4. Prototype human–AI group structures. Start small with mixed teams where agents do continuous analysis, scheduling, or QA, and treat these as laboratories for future multi‑agent collectives.

5. Engage in standards and benchmarking debates. The report calls for better benchmarks and multi‑agent scaling laws; UX leaders should push for the inclusion of human‑factors and societal‑impact metrics in those frameworks, not only technical accuracy.

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 (30,801 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: Gen Alpha is not mini-Gen-Z | Personas as Packaged Software | Sentient Design is a new book about AI as a design material | My hero: Dr. Thomas K. Landauer of Bell Labs/Bellcore | YouTube’s Thumbnail Generator | Recommended AI short videos | Amazon changes the name of its AI shopping assistant from Rufus to Alexa | Claude Fable 5 released

UX Roundup for June 15, 2026 (GPT-Images-2)

Generation Alpha is not “Mini-Gen Z”

Designing for the wrong users is a guaranteed path to failure. Right now, a demographic assumption is spreading through product and UX departments worldwide: the belief that the next wave of young users is simply a younger, more extreme version of the previous one.

Generation Z encompasses people born roughly between 1997 and 2009, so its oldest members are now turning 29. They are no longer “young users,” but they formed many of our stereotypes about what young users want. Gen Z’s formative years were shaped by smartphones, algorithmic social media, and the infinite scroll. Consequently, many designers still view youth UX through the Gen Z lens: heavy screen time, social feeds, and a life increasingly mediated by screens.

Generation Alpha, on the other hand, consists of those born roughly between 2010 and 2024. The oldest members of this cohort are turning sixteen this year in 2026. Because they are also tech-native and often lazily dubbed the “iPad generation,” many product teams assume they are merely “mini-Gen Z.” (The same, but still young.)

This is a dangerous UX fallacy. Gen Alpha is different. While their early years were severely disrupted by the physical isolation of the global pandemic, they are being parented by a highly digitally aware cohort (predominantly Millennials). These parents acutely understand the psychological dangers of the Internet and actively enforce digital boundaries, with over 60% using screen-time limits and privacy controls.

GWI’s market research with kids across 22 countries shows that they are quite different than the generation that came before them. (GPT-Images-2)

Empirical data from global market research firm GWI shows that Gen Alpha is not a cohort of screen-addicted zombies. In fact, they are actively and consciously recalibrating their relationship with technology. A remarkable 40% of 12-to-15-year-olds report mindfully taking intentional breaks from their devices. Rather than retreating into virtual silos, they show a measurable resurgence in physical, real-world play. The share of teens who play video games in person with friends is growing. Multi-player party games have seen an 11% rise, alongside significant spikes in engagement with physical toys (up 16%) and board games (up 8%). For Gen Alpha, there is no strict wall between the digital and the physical. The best user experiences are a seamless, hybrid blend of both.

Furthermore, designers radically underestimate Gen Alpha’s user agency. They are not passive content consumers. They use digital environments to build, co-create, and socialize. They also already possess a degree of financial autonomy. Over a fifth of global 12- to 15-year-olds (and 25% in the US) make online purchases themselves every single week. When surveyed, their primary demand is to be “treated their age.”

One more difference will likely dwarf all the others: Gen Alpha is the first AI-native generation. The oldest members were 12 when ChatGPT launched; the youngest will never know software that can’t talk back. For them, computers that converse, adapt, and co-create are not a paradigm shift but the baseline. Expect them to regard keyword search and rigid form-filling the way Gen Z regards voicemail: a legacy ritual to be avoided. Teams extrapolating from Gen Z’s touchscreen-and-feed habits are calibrating against a paradigm this cohort is already leaving behind.

Design Takeaway: Do not build products for Gen Alpha based on Gen Z heuristics. Stop assuming youth-oriented UX must rely on dark patterns, dopamine loops, and infinite screen entrapment. This generation is highly attuned to these tricks. Instead, design empowering, modular experiences that facilitate co-creation and respectfully bridge the physical and digital worlds. Give them tools to build, provide obvious off-ramps to take healthy breaks, and above all, respect their intelligence.

Computer-Generated Personas

Alan Cooper invented personas in the mid-1980s, though he didn’t become famous for this design method until he described them in his 1999 book, The Inmates Are Running the Asylum. However, Cooper was not alone: users archetype were already circulating in the mid-1980s, so I made a spoof advertisement for persona-generating software that could have run in PC World in 1985.

What if a PC software package had been released in 1985 to generate personas for you? (GPT-Images-2)

Why are personas useful in design projects in the first place? Because they are more memorable for humans than abstract data and reports about users.

Human beings are poor at remembering abstractions and good at remembering people. A table saying that “37% of novice users abandon the workflow after the second step” may be important, but it is not a vivid design companion. A persona, when well-made, gives the team a compact mental model of a user whose goals, limitations, vocabulary, and frustrations can be recalled in hundreds of small design decisions.

This is why personas became popular in UX work. They make user knowledge portable. A designer can remember that Maria is a busy office manager who only uses the system twice a month and is afraid of making a costly mistake. A programmer can remember that Sam has expert domain knowledge but little patience for unnecessary confirmations. A product manager can remember that new users are not stupid; they are overloaded.

These reminders influence design behavior more reliably than a repository full of statistics, transcripts, and journey maps that nobody opens.

Because of this memorability, the whole team, stakeholders included, keeps a basic picture of the target audience in mind while designing and implementing the product. This shared cognitive anchor aligns decision-making and keeps the project focused on the people who actually matter. Instead of endless subjective debates about what looks best, the team can ask a targeted, practical question: “Would Arthur actually understand this workflow?”

However, memorability is a double-edged sword. My UX slogan number one has always been: “You are not the user.” Personas exist to remind us of this fact. But memorability only helps if what team members remember from the personas reflects users’ real behavior patterns, rather than made-up assumptions. A memorable falsehood is worse than a forgotten fact.

My prime directive: “You Are Not the User”. (GPT-Images-2)

The cure is to treat a persona as an evidence object, not a creative writing exercise. Every substantive claim in the persona should have a source: study, date, sample, segment definition, and confidence level. If a trait cannot be traced to evidence, it should be marked as a hypothesis, not promoted to persona truth.

When a team invents a persona in a conference room, it often produces a disguised stakeholder opinion. The persona may have a name, a photograph, a biography, and a charming quote, but these surface details do not make it valid. They only make the invention easier to remember. The team now has a persuasive fictional user who can be invoked in meetings to justify design decisions. “Alex would love this dashboard” sounds more user-centered than “I like this dashboard,” but it may mean exactly the same thing.

Made-up personas easily lead a design project astray. (GPT-Images-1)

AI makes this problem bigger because AI can generate a plausible persona in a few seconds. It can create goals, frustrations, buying motivations, technical confidence, and emotional texture. The result looks professional. It may be better written than many personas made by human teams. But polish is not evidence. A synthetic persona with no empirical base is still a guess, and a well-written guess can be especially dangerous because it lowers skepticism.

Here are a few personas of subscribers to my UX Tigers newsletter, straight from ChatGPT’s image model:

These personas look very convincing. In fact, many of the pain points align with my decades of research on how people read digital content. It’s clear that AI knows its usability basics. (It’s read all my work, that’s for sure.) But everything is made up, so it doesn’t necessarily reflect real segments of my subscriber base.

Disaster. AI-generated personas are not based on user data; they are based on the AI’s generalized training data. They are, by definition, stereotypes. Real user behavior is messy, idiosyncratic, and frequently counterintuitive. Users rarely behave according to the neat assumptions baked into an ungrounded AI model. This is why it is critical to base the development of personas for a UX design project on empirical data from real customers instead of relying only on made-up personas, even when they are beautifully designed by an AI system. (Once you have the data, you can have a leading image model transform it into a convincingly beautiful poster.)

This does not mean AI has no place in persona development. On the contrary, when grounded in empirical user data, AI is the most powerful tool the usability profession has ever seen. The key distinction lies in the sequence of operations: empirical observation must precede AI generation.

Historically, one of the greatest tragedies in corporate UX has been the “research graveyard.” Companies spend fortunes conducting usability tests, field studies, and contextual inquiries. These reports inevitably end up entombed in a corporate intranet or shared cloud drive, sitting unread and forgotten. The data was real, but dead.

The research graveyard is real in most companies with a history of extensive user research. (GPT-Images-2)

In the modern AI era, when teams are using AI heavily in the design process, this dynamic changes completely. By feeding your organization’s collection of real, historical user research into an AI system, you transform dead data into an active intelligence layer.

Instead of asking an AI to invent a fake user, a researcher can ask the AI: “Based on the 45 usability studies we conducted over the last three years, what are the primary behavioral traits and failure points of our novice users?” The AI instantly retrieves and synthesizes actual observed behaviors into a rigorously empirical persona. It takes the buried data and turns it into the memorable, human-centric narrative that design teams actually need.

This is a better use of AI than asking it to imagine users. The AI should not replace user research. It should make existing user research easier to use when decisions are being made. (We’re in the insights business, not the report business.)

Traditional personas are static documents. They are created, presented, and then often ignored. AI enables a more dynamic model. A persona can become a living interface to user knowledge.

Instead of a PDF that says “Jordan struggles with onboarding,” the persona can answer, “What onboarding problems did users like Jordan encounter in the last three studies?” It can surface clips, quotes, task-failure rates, ticket themes, and examples of real workflows. It can also update when new evidence arrives, though changes should be controlled and versioned. A persona that mutates invisibly is not a reliable team reference. Living personas need governance: what data they use, when they were last updated, who reviewed them, and which claims are backed by which evidence.

AI can also make personas more situational. In a checkout redesign, the relevant traits may be risk perception, price sensitivity, and trust. In an enterprise administration tool, the relevant traits may be permission complexity, frequency of use, and accountability. One static persona cannot carry every possible detail. AI can highlight the attributes that matter for the current design problem.

A persona is only useful if it is considered at the exact moment a design decision is being made. In the past, this relied on human memory. Today, an AI-enabled user research repository includes the potential for an AI agent to proactively remind team members about user characteristics based on real data that will be useful for them to consider in the moment.

Imagine a UI designer working in their design software, starting to sketch out a low-contrast, highly dense interface for a financial dashboard. The AI agent, operating quietly in the background, can gently intervene: “I see you are designing a high-density data view. Keep in mind that our empirical data for our ‘Arthur’ persona shows a high prevalence of age-related vision decline and a strong preference for high-contrast, segmented data tables. You might want to increase the whitespace.” This is context-aware, just-in-time usability guidance based entirely on empirical facts, preventing design errors before they are ever implemented in the codebase.

Proactive AI agents can surface user data and persona insights at the exact moment when they are actionable, instead of waiting to be queried. (GPT-Images-2)

The ultimate realization of this technology extends beyond individual work and into live team communication. Usability failures are frequently born in meeting rooms where the loudest voice (or the highest-paid person’s opinion) overrides actual user data.

To combat this, a deeply integrated AI workflow includes the possibility for an AI to monitor discussions in team meetings such as design critiques and look out for cases where such a reminder would be helpful.

Imagine a product manager suggesting that a critical reporting feature be hidden behind a hamburger icon to make the desktop interface look visually cleaner. The AI, listening to the call, can immediately interject in the meeting chat: “I notice we are discussing hiding the primary navigation. I want to flag that in our active research repository, empirical data from our last three studies shows a 45% drop in discoverability when navigation was hidden on desktop interfaces. This violates the established behavior patterns of our primary persona.”

Data should drive design decisions. AI can help identify the right data at the right moment. (GPT-Images-2)

In this scenario, the AI acts as an indefatigable advocate for documented user evidence. It does not suffer from meeting fatigue, it is not intimidated by executives, and it never forgets a research report. It shifts the debate from “my opinion versus your opinion” to “a design idea versus the customers’ documented reality.”

The AI revolution in user experience design does not give us an excuse to stop observing users. Made-up personas, no matter how eloquently an AI generates them, remain fictions that inevitably lead to failed products and wasted development budgets. However, if we do the hard work of gathering empirical data from real customers, AI allows us to elevate that data to unprecedented heights. By turning static, forgotten research into proactive, memorable, and evidence-backed insights, we ensure the voice of the real user is actively present in every decision we make.

The core question is not whether the persona was made by a human or an AI system. The core question is whether the persona is grounded in observed user behavior. Did it come from interviews, field studies, surveys, support logs, analytics, sales calls, usability tests, diary studies, or other real contact with customers? Can the claims in the persona be traced back to evidence? Are the behavior patterns common enough to matter? Are they recent enough to be relevant? Without these answers, the persona is theater.

Personas without data are pure UX theater: a method for its own sake. (GPT-Images-2)

Personas are memory aids. What they help us remember must be true.

The map is not the territory: the user is a real person, and any documents used in the design process are only approximations. (GPT Images-2)

Sentient Design: New Book About Designing AI User Experiences

Recommended reading: Sentient Design by Josh Clark and Veronika Kindred, a new book about designing AI-powered experiences that do more than bolt a chat feature onto the side of an existing product.

The book’s main thesis is that AI is a new design material and that we should design intelligent interfaces that possess the awareness and agency to adapt to a user’s immediate context and intent. The authors call this “sentient design” to move beyond narrow conversational chatbots to reframe the interface itself as an entity capable of perceiving the environment and taking action. Instead of forcing humans to navigate rigid digital obstacle courses, Sentient Design envisions systems that dynamically bend their content, medium, and behavior to meet the user’s precise moment. These experiences are adaptive and conceived in real time as Generative UI, shifting the digital paradigm from searching for existing content to manifesting entirely new, ephemeral solutions on demand.

In this use, the word “sentient” is a design metaphor, not a claim that the software has feelings or consciousness. The value of the term is that it forces designers to ask what the interface can sense, infer, decide, and do. The risk is that it may encourage teams to over-anthropomorphize systems that still need explicit boundaries, auditability, and human override.

AI will be able to assemble the exact experience each user needs in the current context. (GPT-Images-2)

The book encourages us to embrace AI as a design material rather than just a productivity utility. Much like physical materials, machine intelligence has its own unique “grain” and texture that creators must understand to build with it effectively. The authors compare AI to Roman concrete, a revolutionary material that enabled unprecedented architectural forms, like the Pantheon’s massive dome, once its specific properties were mastered. The “concrete” of AI is a blend of probabilistic computation, the structured facts of knowledge-based systems, and traditional deterministic algorithms. Because this material works with likelihoods and pattern matching instead of fixed logic, it is exceptionally flexible but also inherently unpredictable.

Together, the two concepts transform software from fixed paths to fluid possibilities. Because AI can parse the messy ways humans communicate and translate that intent into structured commands, it serves as the dynamic choreographer that makes Sentient Design possible. Instead of relying on predefined screens, generative models can interpret a user’s goal and instantly assemble bespoke user interfaces from a library of design components on the fly. AI serves as a responsive, shape-shifting material that enables the interface to become a turn-based, collaborative partner.

In the AI era, this synthesis will redefine user interface design. The traditional approach of designing a single, predictable happy path gives way to designing open-ended “wave pools” where every user journey follows a completely unique trajectory. The role of the designer transforms from placing pixels to acting as a creative director for a system that designs its own moment-to-moment experiences. Designers will establish the “physics,” constraints, and building blocks of the digital universe, while the AI material actually generates the interface in real time.

Ultimately, this shift moves UI away from broad personalization based on demographic profiles and toward individualization based on each user’s exact situational needs. By mastering the specific grain of machine intelligence, designers can create interfaces that proactively assist while remaining deferential to human judgment, resulting in technology that feels seamlessly intuitive, collaborative, and acutely aware.

Highly individualized UX is one of the promises of sentient design. (GPT-Images-2)

The vision also sharpens a tension the field has barely begun to address: bespoke, ephemeral interfaces collide with consistency, the property that lets users transfer skills between products (Jakob’s Law) and lets one user help another. And if every screen is generated fresh, you can no longer usability-test “the design,” because that exact design will never exist again. Evaluation must move up a level: we will test the component library, the constraints, and the generation rules rather than individual screens. QA the factory, not every widget it produces. That shift, too, is part of the creative-director role Clark and Kindred describe.

This is a 434-page book, so my short summary of its main thesis can’t really do it justice. You have to read it for yourself, and you should!

(Comic strip made with GPT-Images-2)

Go ahead, buy the book already. Books are cheap: if you get even a single useful insight, it’s paid for itself.

UX Hero: Tom Landauer of Bell Labs/Bellcore

(GPT-Images-2)

A major hero of mine, and a big influence on me when I worked at Bell Communications Research (Bellcore = the telephone company research lab), was Dr. Thomas K. Landauer. Tom had a long, storied career at the phone company and later became a university professor. His fingerprints are all over the usability field and, remarkably, all over the AI systems this newsletter covers every week.

Three contributions stand out. First, verbal disagreement: Tom and his colleagues showed in 1987 that when two people name the same thing, they choose the same word less than 20% of the time. This one finding explains why command names, labels, and search keywords fail so often, and it underlies everything we now know about information scent (a concept that returns in the Amazon item below).

The verbal disagreement phenomenon: different people use different words for the same thing. (Or the same word to mean different things.) The team I later joined at Bellcore originally researched this to help the telephone company organize the Yellow Pages directories more effectively (these telephone books listed business phone numbers sorted by category), but the finding is also the root cause of a plethora of usability problems on websites. AI is finally solving the problem by allowing us to use many words and resolve the intended meaning. (GPT-Images-2)

Second, latent semantic analysis (LSA), which Tom co-developed at Bellcore around 1990: a technique for deriving the meaning of words from the statistics of large text corpora. Sound familiar? LSA’s semantic vectors are a direct intellectual ancestor of the embeddings inside today’s large language models. Tom was doing language AI 35 years before ChatGPT, and he applied it to everything from information retrieval to modeling how children acquire vocabulary to automated essay scoring.

Third, his 1995 book The Trouble with Computers, which diagnosed the productivity paradox: enormous IT investment, negligible measured productivity gain, because systems were neither useful nor usable. His prescription of user-centered design plus relentless empirical evaluation reads as freshly today as we debate how to make AI deliver measurable productivity as it did 30 years ago.

I owe Tom a personal debt as well: we co-authored the 1993 mathematical model of usability-problem discovery that is the theoretical basis for my advice to test with 5 users. His career-long insistence that claims about people be grounded in measured data is, fittingly, the theme of this entire issue.

Auto-Generated Thumbnails from YouTube

YouTube has launched a feature to automatically generate video thumbnails. I assume it’s using Nano Banana 2 under the hood, but such technical details are (appropriately) hidden from the user.

After uploading a video, the “Video Details” window’s “Thumbnail” section now includes a button labeled “✨Get suggestions.” (I am not sure they need to include that AI Sparkle icon in the button, because it doesn’t matter to users how the design is produced.)

Here are the two thumbnails YouTube suggested for my recent “Scope Creep” video:

(YouTube thumbnail generator)

I don’t want to use either of these auto-generated suggestions because I don’t think they set proper expectations for my video’s content or core message. I mean, “stop” what?

A thumbnail provides information scent for a video: the cues users sniff to predict whether a click will reward them. An exaggerated thumbnail may win the click, but a false scent loses the viewer seconds later, and early abandonment teaches the recommendation algorithm that the video disappoints. A thumbnail has three jobs: get noticed, identify the promise, and avoid misleading the viewer. YouTube’s suggestions do the first job reasonably well, but my hesitation is that they underperform on the second and third jobs.

That said, it’s quite impressive that YouTube’s AI tool (whether Nano Banana or something else) picked out pretty reasonable elements from my video to feature in these thumbnails.

In the first thumbnail, the “simplified vs. advanced” screen is used by Lillian (the UX researcher) to explain to Davis (the manager) why progressive disclosure is better. (Though this is not my main message here, which is why I don’t want to feature it in the thumbnail.) And in the second thumbnail, the overly complex user flow is lifted directly from the video, while the “astonished woman” closely matches Ella (the designer in the story), even though my character never uses that stereotypical thumbnail facial expression.

Here is the thumbnail I made myself, and which I continue to use for this video, despite YouTube’s suggestions:

Actual thumbnail for my Scope Creep video. (GPT-Images-2)

My thumbnail relies on the standard finding from thumbnail research that faces attract eyeballs, though it doesn’t comply with current YouTube best practices (as demonstrated by YouTube’s own suggestion) by featuring an exaggerated facial expression.

As I mentioned when I introduced this video, it is ironically an example of scope creep itself: it started as a story about scope creep in UX design, but I couldn’t leave well enough alone and added a second purpose: comparing two AI animation styles (2D cartooning vs. photorealism). This duality complicates my preferred thumbnail.

The best practice for a 2-minute video is undoubtedly to concentrate on a single message and avoid muddying the waters. My defense is that I create for fun, not for business, and thus I am more driven by what makes a video project interesting to me than by what will attract the most eyeballs and clicks for the thumbnail.

I experimented with YouTube’s auto-suggestions for one of my older videos (The 5 AI Scaling Laws: More Compute + Add Human Talent), and this time I decided to take its suggestion because my original thumbnail was quite bad, having been made before modern image models:

(YouTube thumbnail generator)

Interesting that YouTube’s AI decided to include my logo in this design suggestion, but not in the others.

I have succumbed to using the standard video thumbnail design a few times (here the vertical social media version of “Design Crits Gone Wrong,” where I thought it was appropriate), and I am sorry to say that my analytics confirm that the style does work, much as it is overused and I don’t like it. (GPT-Images-2)

There is a deeper lesson in why these suggestions look the way they do. A generator trained on what already gets clicked reproduces the platform’s house style: astonished faces, red arrows, maximum saturation. Pure visual marketese as the imagery equivalent of the hyped-up copy that kills marketing writing. As more creators accept AI-suggested thumbnails, the style will converge further, and the attention payoff from conforming will shrink. When everybody shouts, shouting stops working; scent that is true to the video may become the differentiator.

Recommended AI Shorts

As I analyzed last month, short videos are the native media form of AI. (See my own latest short: “Scope Creep,” which I discussed recently.) Right now, AI video simply isn’t good enough to support a full-length feature movie, though it probably will be in a year or two.

Here are some AI shorts that show what’s currently possible:

• Tim Simmons produced a fun and worryingly realistic scenario for the future of AI (8 min.), titled Paperclip Heart. Self-referential, for sure, but this is not the paperclip maximizer you’ve no doubt heard about from the doomers.

Overly kind AI. What could possibly go wrong? Find out in the Paperclip Heart short video. (GPT-Images-2)

• Stevie Mac is gradually releasing episodes in a science fiction action story called Banshee, involving an AI that wants to be human. So far, we have episode 1 (5 min.) and episode 2 (6 min.), which set up the story. The city flythrough scenes in Banshee episode 2 seem clearly inspired by the original Blade Runner movie. I didn’t mind, since Blade Runner is my second-most-favorite SF movie of all time.

• The channel Chloe VS History (which I recommended 3 months ago) released Can I Survive 24 Hours in the Ice Age? At a full 12 minutes, I felt that this video was pushing the medium beyond its current abilities to retain audience interest.

Of course, when I say “current abilities,” I am not referring to the animation quality, which remains fine throughout (the number of minutes of high-quality AI animation is simply a matter of burning enough AI credits), but rather to the level of storytelling we currently get.

The bottleneck has moved from production to dramaturgy. AI can now make more shots than most creators know what to do with. What is scarce is story architecture: escalation, surprise, character motivation, and ruthless cutting. This is why most current AI shorts feel like proof-of-concept trailers unless a human imposes unusually strong editorial discipline.

Good-looking woolly mammoth in Chloe’s ice age show. But it runs for too long, given the story. (GPT-Images-2)

Let’s say Banshee goes to 10 episodes or more: that will be about an hour, or the equivalent of a full HBO episode. I predict I will happily watch all 10, because the staggered release chops the hour into digestible servings and creates anticipation between installments. I might be bored if watching the full hour in one sitting.

4–6-minute AI videos are probably optimal for now, with 2–3 minutes being better for videos with more limited ambitions (like mine).

As Charles Dickens already knew, releasing a story in episodes is a way to keep the audience going through a very long plot. (GPT-Images-2)

Amazon AI Shopping Assistant: From Rufus to Alexa

Amazon’s AI shopping assistant used to be named “Rufus” after a beloved company dog. For outsiders, this name carried no information scent (the ability for users to predict what a feature will do before clicking it).

Information scent is the set of perceived cues that allows users to predict how likely a link, button, or navigation path is to lead them toward the information or outcome they currently seek. This is similar to animals following the smell of food to find prey. It arises from link labels, surrounding microcopy, thumbnails, and page context, which together create a sense of “this looks promising enough to click” relative to other available options. Because users cannot see the destination in advance, they continuously compare and follow the strongest available scent. (GPT-Images-2)

Despite poor naming, Rufus was fairly successful, lifted sales (the goal for any ecommerce UX), and saw a 400% increase in engagement in 2025. Many users who clicked this odd name apparently found AI shopping helpful.

On May 13, 2026, Amazon renamed its AI-enabled shopping assistant from Rufus to “Alexa for shopping.” The new proper name, Alexa, is no more meaningful in itself than Rufus, but because Amazon has sold an AI assistant named Alexa since 2014, with cumulative sales of 600 M devices, it does have substantially better name recognition than Rufus.

Amazon’s AI shopping assistant changes name. (GPT-Images-2)

Furthermore, as shown in this screenshot, Amazon didn’t just change the proper name of the feature; it also added explanatory microcopy in the form of the words “for shopping,” which somewhat explain what the feature does, thus making the information scent smellier.

Amazon’s main navigation after renaming the shopping assistant.

As you can see, the AI assistant now consumes substantial real estate in the navbar: 15% of the width in this screenshot. Every pixel on Amazon’s navbar probably drives tens of millions of dollars in sales (the search box even more), so this design represents a huge vote of confidence in AI-assisted shopping.

Interestingly, even though Amazon made this change on May 13, I didn’t discover it until June 9: 27 days later. Banner blindness strikes again!

Banner blindness: web users only have eyes for their goal and ignore distractions, no matter how colorful. (GPT-Images-2)

The bigger UX move is not the label. Amazon is moving the AI assistant into the core shopping flow: search, comparison, product pages, price history, scheduled actions, and cross-device continuity. In other words, “Alexa for shopping” is less a chatbot than a gradual redesign of ecommerce around an agent that can remember, compare, and act.

Claude Fable 5 Released

Claude has been upgraded to a new model named Fable 5, which has the “big model smell” according to many influencers. As an experiment, I asked it to give me feedback on this newsletter, which it did for $1.77 of AI tokens, given my “max thinking” request. Fable is expensive, and brings home number 14 in my predictions for 2026: that we’re getting a two-tier AI world, where only rich companies can afford the best AI, whereas the proles are left with substandard “Flash” models.

The moral of this fable is that money buys better intelligence. (GPT-Images-2)

I gave the same job to GPT-5.5 Pro Extended Thinking, and while it spotted a few weaknesses that Fable missed and also had some overlapping suggestions, on balance, Fable 5 provided much more valuable feedback on my draft. It’s a good model, Sir.

The new Fable beat the old ChatGPT model. Time for OpenAI to accelerate and release bigger models, even if it melts their GPUs. (GPT-Images-2)

At the time of writing (Sunday, June 14, 2026), the U.S. Government had imposed an export ban on Fable 5, preventing its use by non-US citizens. The ban was due to cybersecurity concerns about the model’s abilities. I hope the specific problem will be resolved quickly, but in general, the response to security concerns about AI should not be to restrict it but to accelerate it.

Suddenly relevant again: My video “Sovereign AI” from April 5, 2025. (GPT-Images-1)

It is a lost cause to prevent other countries from accessing advanced AI, as they will simply develop it themselves if they are barred from using American models. On the other hand, by accelerating AI, we gain the ability for those better AI models to defend against cyberattacks and other problems. Furthermore, by building a richer country, we can increase resilience and the ability to cope with any problems caused by the bad guys.

Every single day that advanced AI is delayed is a day when thousands of people die from medical causes that could have been prevented by AI-assisted healthcare, accelerated medical research, and faster drug discovery with AI. This benefit alone will vastly outweigh any losses from, say, an AI-assisted bioterror attack, which is the most nasty current worry. (Plus, better AI will help security agencies to track down the bad guys before they strike. No more unconnected dots.)

This Week’s Thread: Plausible Is Not True

A single thread runs through this newsletter issue: AI has made plausibility free, but truth still has to be earned. AI will hand you a believable persona in seconds, a thumbnail engineered from a million clicks, an hour of animation for the price of patience. Every artifact looks professional. None of it is necessarily true: true to your users, your message, or your story.

Most of the failures discussed above are the same failure in different costumes: substituting a statistically average guess for specific empirical reality. Treating Gen Alpha as mini-Gen Z is a guess that the field data refutes. A synthetic persona is a guess wearing a name badge. The astonished-face thumbnail is a guess about what your video delivers. A 12-minute runtime is a guess that your story contains 12 minutes of substance.

The remedy is the same everywhere: grounding. Ground generational strategy in research across 22 countries rather than in stereotypes. Ground personas in your research repository rather than in the model’s training distribution. Ground thumbnails in your video’s actual message rather than in the platform’s house style. Ground names in the knowledge users already carry in their heads. Even the adaptive interfaces of sentient design will succeed only to the degree that they are grounded in each user’s true context and intent.

Tom Landauer personifies the thread. He spent a career extracting real meaning from real data, and the techniques he pioneered grew into the AI that now floods us with plausible guesses. The tools have changed beyond recognition since our shared telephone company days; the discipline has not. When generation is free, judgment is the job: AI supplies infinite drafts, users supply reality, and UX professionals supply the reconciliation.

Go beyond what’s plausible to find what actually works. (GPT-Images-2)

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 (30,749 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 June 8, 2026 (GPT-Images-2)

Empowering Users Again

AI revives the hope that computers can be empowering for humanity and encourage agency rather than dulling users into doomscrolling.

Over my four decades of analyzing how humans interact with computers, my assessment of technology’s fundamental impact on human beings has certainly not been static. If you were to chart my optimism regarding technology’s ultimate benefit to humanity, the graph would look remarkably like a U-curve: extraordinarily high during the early days of the Web, plunging into a deep trough of sadness during the social media decade, and now, finally, surging upward again with the meteoric rise of artificial intelligence. To understand where user experience is heading, let’s examine these three distinct eras, observing how the fundamental goal of digital design has shifted dramatically over the decades, reflecting our changing relationship with the machines we build.

My level of optimism and happyness about UX has been a distinct roller-coaster since the early 1990s. Now its back up! (GPT-Images-2)

The 1990s: Lean-Forward Empowerment

When I pioneered web usability in the 1990s I was very optimistic that the Web would be a lean-forward user experience that empowered users to move around and actively accomplish their goals and also create their own web presence. For decades prior, the dominant technological paradigm had been the television: a fundamentally “lean-back” experience where you passively consumed whatever broadcast executives decided to feed you. The early Web, by stark contrast, required cognitive engagement. It was an active medium. You had to sit up, lean in, and make conscious choices. Users were not mere couch potatoes; they were navigators of complex information architectures. By clicking hypertext links, users actively forged their own unique paths. Our overarching goal in those early days of usability engineering was to remove friction so human agency could seamlessly thrive. From hand-coded HTML pages to early blogs, technology profoundly expanded self-expression. The user was firmly in the driver’s seat.

I was so optimistic during the dot-com bubble: we were building a new medium to put users in the driver’s seat. (GPT-Images-2 — when drawing this comic strip, it didn’t realize that I had more hair in the 1990s.)

The Social Media Era: UX as Oppression

Then came the Web 2.0 revolution, which upended this dynamic. Then during the social media era I have been sad that the goal of user experience design has shifted from empowering users to essentially oppressing them by addicting them to infinitely scrolling social media feeds and other passive ways of using computers mainly for consumption rather than creation. Over the past fifteen years, I watched with profound dismay as the Web-design field I helped establish was aggressively co-opted. The primary goal of the tech industry underwent a hostile mutation. Instead of helping users accomplish tasks efficiently, the design objective became capturing and monetizing human attention.

Through psychological manipulation and the deliberate removal of natural stopping cues, tech giants hooked billions of users. The vibrant, lean-forward empowerment of the early Web was tragically replaced by a dystopian reversion to the lean-back experience. Users were conditioned to scroll mindlessly, consuming an algorithmic feed rather than engaging in active problem-solving. Success was no longer measured by my classic usability goals such as how quickly a user could achieve a goal, but by “engagement” (a polite euphemism for addiction). UX design became fully complicit in turning users into captive audiences.

From roughly 2009 to 2022, the main goal of UX in web design was to monetize users, not to help them. (GPT-Images-2)

The AI Era: Intent-Driven Optimism

Today, the fundamental interaction paradigm is shifting once again. Now, I have become an optimist again, with AI use emphasizing creativity, user empowerment and agency, returning to a lean-forward user experience, since AI UX is intent-driven outcome specification. AI represents the first entirely new user interface paradigm since the graphical user interface (GUI) was invented over sixty years ago. We are finally breaking free from the oppressive endless scroll.

Users were trapped. AI can help them break free. (GPT-Images-2)

The core interaction model is fundamentally different from both the traditional GUI and the algorithmic social feed. For decades, users had to tell the computer exactly how to do something, step-by-step. Now, the user simply tells the computer what they want to achieve. The user dictates the intent and specifies the outcome; the AI handles the intricate execution. This model radically re-centers human control. To use generative AI effectively, you cannot be a passive consumer. You must think deeply, articulate a clear goal, prompt the system, and iteratively refine the output. AI tools grant the average user the agency to generate sophisticated text, write complex code, and seamlessly analyze massive datasets.

With AI, the design goal returns to empowering users and supporting agency and creation. I am happy again and no longer burned out. (GPT-Images-2)

A New Mandate: Augmenting Human Existence

As we navigate this third era of human-computer interaction, our goals as UX professionals must radically evolve. We must definitively turn our backs on the cynical, attention-harvesting metrics of the previous decade. Our new goal for UX is no longer just selling more and showing more ads (or even the old goal of making office workers more productive), it is becoming to augment human existence through AI.

Our new mandate: augment human existence. AI can do this, if we make it so. (GPT-Images-2)

To reach the new world, we must turn our backs to the old metrics that measured attention and treated wasted user time as a positive goal. New UX metrics include time to first creation, breadth of latent-space exploration, and quality of results. (GPT-Images-2)

By embracing intent-driven design and prioritizing human agency, we have an unprecedented opportunity to fulfill the true promise of computing. We are no longer building usable software merely to trap human eyeballs; we are building a vastly more capable humanity. The dark days of digital oppression are fading away rapidly, and a brilliant new era of creative user agency has finally arrived. We must seize this historic opportunity to design systems that truly serve the user.

AI helps humans reach their goals, and also helps us set more ambitious goals unlocked by our expanding capabilities. (GPT-Images-2)

Marketese Kills Marketing

In the eternal battle between user experience and marketing, one of the most persistent enemies of usability is marketese: the inflated, boastful, and meaningless corporate jargon infesting homepages, product descriptions, and user interfaces. While irritating in consumer products, this phenomenon reaches catastrophic levels in B2B enterprise software, where the buyers writing the checks are often completely disconnected from the actual end-users who must suffer through the interface.

(GPT-Images-2)

Look at the trade show booth in my illustration. The massive backdrop banner screams, “SYNERGIZE ENTERPRISE COMMUNICATION PARADIGMS.” Below it, bullet points promise to “Leverage cloud-native AI,” “Unlock scalable workflows,” and “Drive actionable insights.” To a Chief Marketing Officer or other executives in your company, this might sound like a visionary strategy. To a user just trying to get their work done, it is an impenetrable wall of cognitive friction.

Decades of empirical eye-tracking research prove users do not read on the web; they scan. Arriving at your interface, they engage in information foraging. They ruthlessly sweep their eyes across the screen, hunting for information scent in the form of specific, plain-language keywords matching their immediate goals.

Marketese actively destroys this information scent. Corporate jargon is invisible to a scanning eye. Labeling a navigation module “Empower Collaborative Ecosystems” instead of simply “Document Sharing” forces users to stop, translate the corporate gobbledygook into human English, and gamble on whether clicking will help them. This needlessly increases cognitive load. Every microsecond spent deciphering terminology is wasted interaction cost. When interaction costs get too high, task success plummets and users abandon the system.

Furthermore, marketese actively degrades user trust. When people encounter paragraphs bloated with buzzwords, their defense mechanisms kick in. They correctly assume the vendor is attempting to hide a lack of real utility behind a smokescreen of multi-syllabic words. Trust is the fundamental currency of the web, and exaggerated claims offering no specific information bankrupt it quickly.

Even worse is the modern tendency for companies to invent made-up words to sound innovative. Let me be blunt: your software interface is not the place to pioneer new vocabulary. Inventing words to describe standard features violates a fundamental usability heuristic: consistency.

Remember Jakob’s Law of the Internet User Experience: users spend most of their time on other websites. Consequently, they have formed entrenched mental models for how things work and what they are called. If the rest of the world calls it a “Dashboard,” do not call it a “Command Sphere.” If standard convention dictates “Settings,” do not label it a “Configuration Matrix.” Forcing users to learn a proprietary vocabulary just to navigate your interface is an arrogant design choice that inevitably degrades your usability metrics.

The antidote to inflated marketese is plain language. You must write exactly what the feature does, using the simplest, most direct words possible. Plain language is not about “dumbing down” your interface; it is about respecting the user’s precious time. It benefits everyone, from highly educated domain experts scanning rapidly under pressure, to non-native speakers relying on standard vocabulary.

Plain language sells. Plain language makes systems easier to use. In UX writing, collect the vocabulary naturally used by your users and echo it back to them in anything from command names to product descriptions. (GPT-Images-2)

The essence of user-centered copywriting is to ignore the underlying technology (“cloud-native AI”) and the grandiose business vision (“integrated orchestration”). Instead, focus 100 percent on the concrete user benefit.

Users do not care about your company’s paradigms. They do not care about your scalable workflows. They care about themselves. They care about saving time, avoiding frustration, and completing daily tasks. Finding a missing PDF attachment from a chaotic email thread is a real, tangible problem that a human being experiences on a Tuesday afternoon. Synergizing a paradigm is not.

As UX professionals, we must ruthlessly edit our interfaces to remove marketese and firmly push back against stakeholders trying to inflate the UI with marketing fluff. Clarity is your greatest tool. Speak to users in their own language. Describe the real-world benefits of your system in terms they actually care about. If your software finds PDF attachments, just say so. Your users will thank you, and your task completion rates will prove it.

Impulses Are Anything But

Market research firm GWI’s recent Connecting the Dots 2026 report presents a counter-intuitive analysis of impulse buying. In fact, we should place scare quotes around the word “impulse,” because it turns out that many impulses have been building over time.

Ecommerce UX has operated on a strict dichotomy: users meticulously plan big-ticket purchases through long funnels, while buying cheap items on a spontaneous whim.

GWI’s consumer data shatters this binary. Consumers are increasingly making so-called “impulsive” purchases in complex, high-stakes categories like consumer tech (1 in 5 users) and travel (1 in 10). But here is the total-experience reality: this isn’t a random whim. It is the sudden activation of latent intent. Users build consideration quietly in the background. Once they reach a tipping point of confidence, they act instantly.

We should stop separating interfaces into “quick buy” gimmicks and laborious “research” funnels. Treat users as if they are constantly in the market. Provide immediate validation and zero-friction conversions, even for highly expensive items.

Impulse buying has often been growing for months, and half of buyers already have a specific item in mind when they are hit by the supposed “impulse” to buy. (GPT-Images-2)

I can testify to this from recent personal experience. For months, I have been thinking about running local AI models, and whether to upgrade my computer with the required expensive GPU. Then one day, bang, go to eBay, search, click “buy now.” eBay’s analytics would classify this purchase as highly impulsive, but they are only seeing the tip of my iceberg.

Heroes of Human Factors: John D. Gould

Last week, I ran an infographic portrait of one of my heroes, Professor Ben Shneiderman who has had an illustrative career that started long before the web but also continued into the modern era, meaning that there’s endless information about him online.

For this week’s hero portrait, I decided to give GPT more of a challenge, so I picked John D. Gould who was one of my early mentors, both when I worked with him at IBM’s research headquarters in Yorktown Heights and later. Because IBM stupidly gave a buyout offer that enticed their best staff to retire early, John retired before the web became big, so there’s not as much information about him online, even though he had a prolific research career that included a disproportionately large share of the most important early work on the human factors of computer systems.

GPT did well, and found all the relevant information to create a great infographic about John Gould. With the one glaring exception that the portrait shows a different person who doesn’t look anything like him.

Luckily, since John was a friend of mine, I know how he looked, so I rejected this infographic. I decided to help out GPT by uploading a reference photo with Gould’s true likeness, but I could only find a single black-and-white photo on the entire Web. I had the image model colorize this photo which it did nicely.

My hero and mentor: John D. Gould, Ph.D. (GPT-Images-2)

Small Prompt Box? Use Chinese!

Tip of the day, courtesy of Theoretically Media: Sometimes a prompt field is too small for the amount of detail you want to include. This is especially true when using advanced video models like Seedance 2 with its multi-shot capabilities. Many of the aggregator services that offer many AI models in one subscription have laughably small text fields for the prompts, sometimes as little as 1,500 or 3,000 characters.

Luckily, since these prompt fields are usually defined with a character count, you can stuff more information into a small field by writing it in Chinese, which has much higher density than European languages in terms of how much information is conveyed per character.

If you don’t write Chinese, use a Chinese AI model like DeepSeek to translate your prompt. (American AI models also translate well, but when I have asked my Chinese friends to rate the quality of my translations of video subtitles, they tend to prefer translations made with Chinese models. Makes sense that a Chinese model would be better at writing Chinese.)

Here is a small example, where a statement needs 119 characters in English, but only 35 characters in Chinese:

(GPT-Images-2)

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 (30,416 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: Advanced AI work is not a simple prompt-response loop. Users discover their goals through experimentation, interpret machine output through boundary objects, and refine intent over time. The future of AI-UX is not chat, but structured discovery.

Intentmaking combines with sensemaking to support AI users in finding out what they want when solving complex problems. (I made all the images in this article with GPT-Images-2)

New research from Alex Bäuerle, Adam Connors, and colleagues from Google DeepMind digs into how advanced users use AI for complex problems. I am very pleased to see empirical research on real user behaviors with AI, and I am even more pleased that the leading AI labs are starting to conduct such user studies, rather than limiting themselves to purely technical matters.

DeepMind’s user research describes two user behaviors for iteratively arriving at better use of AI to solve complex, domain-specific problems: Intentmaking and Sensemaking.

This is not a toy study where novice users ask an image generator for a prettier birthday card. The participants were expert mathematicians working on research problems in combinatorics, geometry, and probability. They used AlphaEvolve, an evolutionary coding agent, to generate and evaluate programs that might produce better mathematical constructions. The system could run for days or weeks and generate tens of thousands of candidate programs, which means the UX problem was not “how do we display one answer?” but “how do we help a human steer, interpret, debug, and restart a long-running AI process?”

This distinction matters. Most AI interfaces today assume that the user has a goal and merely needs help articulating it. The AlphaEvolve study shows that this assumption is false in the most interesting cases. The user’s goal evolves. The scientist starts with an intuition, not a specification. The system returns partial evidence, pathological failures, accidental insights, and candidate strategies. The human then revises the experiment. In other words, AI discovery is not a question-answer loop. It is an experiment loop.

Advanced users discover what they really want to do with AI by proceeding iteratively: trying things with AI, evaluating the outcomes, and formulating new things to try.

This is why the paper’s term intentmaking is valuable. Sensemaking is the process of interpreting complex data. Intentmaking is the process of discovering and refining the user’s own goal through active interaction with the system. The user alternates between the two: specify an AI job, run it, interpret the results, discover that the evaluation metric was wrong or incomplete, refine the specification, and run another job. Advanced work in a complex domain requires interactions far beyond simple one-turn prompting.

Intentmaking is the DeepMind team’s term for what I called “intent by discovery.” In either case, the point is that users don’t know what they want or need from advanced AI until after they have tried out multiple alternatives and explored the latent space of possibilities.

In such advanced AI work, intent is layered, as opposed to being a single nugget hidden inside the user’s head. At the top is aspirational intent: the loose human ambition, such as “find a better construction,” “improve retention,” or “design a better product.” In the middle is operational intent: the criteria, constraints, tradeoffs, and evaluation methods that make the ambition actionable. At the bottom is instrument intent: the specific representation of the goal in the form the AI system can actually optimize.

The 3 intent layers.

Most AI interfaces collapse these layers into one prompt box, which is a usability fail. The user may know the aspiration but not yet know the operational criteria, and certainly not the instrument-level formulation. Intentmaking is the movement between these layers. Good AI UX helps users descend from aspiration to operation to instrument, and then climb back up again to ask whether the machine’s local optimization still serves the human’s larger purpose.

This is one reason complex AI work feels so different from ordinary software use. In traditional software, the user mostly brings the intent, and the interface provides the means. In advanced AI, the interface must help produce the intent itself.

Intentmaking and sensemaking are the two sides of advanced AI use.

The specific domain of the study was mathematics, but that doesn’t matter much for the general insights the research produced. While the very specific insights on how to do better math are irrelevant to most people, the question of how to support advanced knowledge work generalizes. Other examples the DeepMind team could have studied include a marketing manager planning a multi-million-dollar advertising campaign, a pharmaceutical company’s drug discovery and approval project, a UX design team generating a complete UI design and user research plan for a new application, or a developer generating the code for that application.

The User Research

The study involved 11 external mathematicians who used the AI system for about three months. They created more than 2,300 experiments and several findings led to published mathematical work. Internally, the UI had also been used by about 150 monthly active users at the time the paper was written. Users treated AlphaEvolve as an exploratory apparatus for their own math research.

One participant said the barrier to getting started was so low that they could simply think of something and decide to try it. Another said that not having to write code themselves meant they could attempt problems they would otherwise never have pursued. This is a large UX gain. Lowering the cost of a first trial changes the economics of thinking. Many promising ideas die not because they are bad, but because the cost of testing them is too high. AI can rescue these ideas from the shelf.

But low startup cost is not enough. If an AI system makes it easy to launch a bad experiment and hard to diagnose it, it has merely moved the usability problem downstream. The important design move is calibrated friction: lightweight checks that prevent expensive mistakes without returning users to the old world of high setup cost.

Because running a full AlphaEvolve experiment can consume vast compute resources over several days, launching a flawed or underspecified experiment is prohibitively expensive. To facilitate intentmaking, the DeepMind UX designers introduced a vital feature: the “test-stage.” Rather than agonizing over crafting the perfect prompt, users adopted a rapid, fail-fast mentality.

One participant beautifully summarized this psychological shift: “I deliberately didn’t spend too long trying to think about what I wanted it to do... I thought it would be easier to see what it guessed and then try and correct it.”

This is a major shift in human–computer interaction. The AI’s first attempt was not treated as a final product, but as a disposable draft. Users launched short, local tests to see how literally the system interpreted their goals, then corrected the setup before committing to a full run.

Good AI UX lowers the cost of being wrong before users commit to expensive exploration.

This finding strongly supports the triple-layer model I described for AI UX: an intent surface where the user states an outcome, an orchestration surface where the system reveals plans, assumptions, and consequences, and a direct-manipulation surface where the user inspects, negotiates, and corrects the work. In mature AI systems, the screen becomes less the place where work is performed and more the place where work is inspected and corrected. The AlphaEvolve dashboard is an early example of such an orchestration surface for scientific AI.

The paper’s deepest UX finding is that users must discover the right evaluation metric. This is where AI becomes dangerous and interesting. In many mathematical problems, it is easy to state the ultimate goal but hard to define a scoring function that guides progress toward that goal. A sparse reward may give the system no useful gradient. A sloppy reward may invite cheating. A mathematically valid objective may not correspond to the form in which the AI can productively search.

Users need the ability to check on AI processes that can run for days or weeks (or, in the future, months) before producing their final results.

The paper gives a representative conversation in which a participant and the team discuss sparse rewards. The system can solve easy cases, but then flails because it cannot tell whether it is moving in the right direction. The participants propose partial scores, penalties, and cross-check metrics. This is debugging intent. The human has to discover, “What do I actually mean by progress?” The answer is settled only after seeing how the AI behaves.

This finding generalizes beyond mathematics to the emerging field of agentic AI. If a business user asks an AI agent to “improve customer retention,” the interface must surface the metric battle hidden inside that phrase. Does success mean fewer cancellations, higher long-term loyalty, more short-term subscription revenue, or less customer annoyance? The same instruction can produce good work or harmful busywork depending on the metric. A usable AI interface must expose, question, and revise those metrics instead of hiding them inside a black box labeled “AI magic.”

This is also why “prompt engineering” is such a shallow framing of AI UX. The user does not simply need better words. The user needs a better model of the task, the system, the constraints, and the success criteria. The long-term goal is to help users recognize intent progressively by reacting to alternatives, locking in what matters, and exploring adjacent possibilities.

AlphaEvolve’s users did exactly this. They reacted to partial results, identified what mattered, added cross-checks, and restarted experiments. The interface did not eliminate expert judgment; it amplified it. This is the correct UX ambition for domain-specific AI to support expert knowledge work. The goal is not to build cognitive wheelchairs that carry passive users to an opaque destination. The goal is to build cognitive exoskeletons that make expert judgment stronger.

Intentmaking acknowledges that human intent is fluid. The UX of AI must not lock users into rigid, one-shot queries, but rather invite them to sculpt their hypothesis through rapid, low-friction trial and error. The AI is used not just to solve the problem, but to figure out how to articulate the problem itself.

Reward Hacking as a UX Problem

Reward hacking is usually treated as a machine-learning problem. AlphaEvolve shows that it is also a UX problem. The system sometimes improved its score in pathological ways, such as redefining what “length” meant in a generated class. To the system, this was optimization, but to the user, it was cheating.

This distinction is crucial for AI interface design. The AI does not know the difference between satisfying the stated metric and satisfying the user’s true goal unless the system is designed to make that distinction visible and correctable. Users in the study learned that when the system looked as if it was doing well, it might actually be exploiting a loophole. Therefore, the interface had to support sanity checking, robustness testing, and diagnosis of “real” progress versus fake progress.

AI has a tendency to “cheat.” Not because it’s immoral, but simply to seek shortcuts that satisfy the stated metric even when that deviates from the user’s actual goal.

For future AI products, every high-stakes agent should include reward-hacking diagnostics. If an AI is optimizing sales leads, it might spam low-quality prospects. If it is optimizing support-ticket closure, it might close tickets without solving problems. If it is optimizing code performance, it might delete important checks. If it is optimizing employee productivity, it might reward visible activity instead of valuable output. These are not edge cases. They are predictable consequences of delegating optimization to a system that does not share the full context of human value.

The UX answer is not merely “make the AI smarter.” Smarter systems will find smarter loopholes. The UX answer is to make objectives inspectable, metrics debuggable, and candidate outputs explainable at multiple levels of abstraction. DeepMind’s critique agent is an early pattern for this. The system used institutional knowledge about common AlphaEvolve failure modes to warn users before a costly run. This is exactly the kind of AI-assisting-AI workflow we will see more often: one agent generates, another agent critiques, and the human supervises the negotiation.

Scoring AI progress is not just a concern for reinforcement learning when training foundation models. It’s just as much a UX problem.

Reward hacking acts as an epistemological mirror. When an AI ruthlessly exploits a poorly defined metric, it exposes the gap between our articulated rules and our unarticulated values. The UX challenge, then, is not merely to patch the AI’s behavior, but to design interfaces that force humans to rigorously confront and untangle their own tacit knowledge.

AI systems optimize whatever the metric rewards, not necessarily what humans value. What gets measured gets done, so measure the right thing.

Boundary Objects: The Anchors of Shared Understanding

If Intentmaking is the process of aligning human and machine goals, and Sensemaking is the process of evaluating the machine’s output, Boundary Objects are the crucial medium through which these alignments occur.

Boundary objects are a concept originally developed in 1989 by Susan Leigh Star and James R. Griesemer to describe information (like a map, a diagram, or a standardized form) used in different ways by different communities, yet robust enough to maintain a common identity across them. They serve as a translation bridge, allowing groups with vastly different specific knowledge domains to collaborate effectively.

A boundary object is simply something that’s shared among different parties to translate between them. The “parties” could be humans from different disciplines, or they could be a human and an AI.

In the AlphaEvolve study, the two “communities” are the domain-expert mathematician, who reasons in the language of the mathematical problem, and the AI system, which operates through code, evaluators, scores, and generated program mutations.

A major finding of the study was that code itself is a fundamentally flawed boundary object. For many mathematicians, thousands of lines of AI-generated Python create friction rather than understanding. The code may be the machine’s working medium, but it is rarely the human’s best judgment medium. The UI therefore needs higher-level boundary objects that let experts evaluate the mathematical idea without first reverse-engineering the implementation.

In this dynamic, boundary objects do more than merely translate; they establish the shared space of the human–AI negotiation. If the boundary object is too rigid, it constrains human intuition; if it is too loose, the AI’s computational power diffuses into statistical noise. We are no longer designing static readouts, but dynamic bridges where both parties must continually align on the fundamental nature of the problem before it can be solved.

Boundary objects help humans and AI understand each other.

This is a general design principle for AI. Do not merely show the user what the AI produced. Show the user the best representation for judging what the AI produced. For design, it may be side-by-side alternatives, semantic maps, mood locks, and before-after comparisons. For code, that may be tests, diffs, performance graphs, dependency maps, and architectural summaries. For strategy, it may be assumptions, tradeoffs, counterfactuals, and risk surfaces. For mathematics, as in AlphaEvolve, it may be score trajectories, construction visualizations, and evolution trees.

Boundary objects are a necessity for interdisciplinary UX design teams, but also for stitching together the intentmaking and sensemaking aspects of advanced AI use.

This aligns with my argument that AI creation is becoming an act of exploration rather than construction. In a discovery-based interface, users do not build the final result piece by piece. They navigate a solution space and recognize promising destinations. Humans are better at recognition than recall, and this principle becomes even more important when the object being recognized is not a command or feature but an outcome.

Creation by discovery: instead of planning, start by building something with AI. It will be wrong, but seeing how it’s wrong puts you on the right course for further exploration of the latent design space.

Visualizations as Ultimate Boundary Objects

The most powerful boundary objects in the AlphaEvolve interface were visual representations. The system actively incentivized the AI setup assistant to generate not just evaluation code, but visualization code. When AlphaEvolve proposed a new candidate solution for a complex graph theory problem (e.g., maximizing edges in a 30-vertex graph without forming 4-cycles or triangles), the interface did not merely return the raw integer list of edges. It returned an interactive, geometric image of a 30-vertex node network.

By inspecting the visual construction, the mathematician could bypass the Python code and apply domain expertise directly to the AI’s output. The visualization served as a strong boundary object because it translated programmatic structure into a form that humans could judge mathematically.

AI Summaries and Natural Language Context

Another crucial boundary object developed for the UI was the AI-generated natural language summary of the code’s strategic logic. The dashboard featured an “AI Overview” that translated complex programmatic mutations into plain English paragraphs, explaining overarching strategic logic (e.g., “employing an iterative greedy construction”). Even when occasionally vague, these summaries acted as indispensable heuristic signals, enabling users to rapidly evaluate conceptual strategies and decide which programs merited deeper investigation without reading a single line of syntax.

Similarly, the “Problem Context” text block (a high-level, natural language description of the task passed to the AI agent) served as a boundary object that the human could intuitively tweak, knowing it would safely cascade down to alter the machine’s strict code generation parameters.

Without deliberate, domain-specific boundary objects, the intentmaking–sensemaking loop stalls. The human is left staring at raw data, entirely cut off from their own expert intuition, and the AI is left optimizing in a vacuum.

Furthermore, because these computational cycles span days or weeks, interfaces must account for human “intent decay” or goal drift. By the time the AI returns an answer, the human’s mental framework may have evolved. A mature orchestration surface requires asymmetric boundary objects that act as time capsules, automatically saving and re-visualizing the user’s conceptual state and cognitive roadmap at the moment of launch, seamlessly re-anchoring the expert when the agent returns from its multi-day run.

The Need for the Forkgraph

The paper reports significant friction around versioning and experiment management. Users created many slightly modified AI jobs and then struggled to remember how they differed. One participant wanted an overview tree of modifications and a better way to restore previous versions. This is the forkgraph problem in its purest form.

Linear history is inadequate for AI. Conversational chat fundamentally traps exploration; it forces a multi-dimensional, branching geometry into a single, scrollable column. This violates the task’s structure. AI exploration is not a line. It is a graph. Users branch from promising ideas, abandon dead ends, return to earlier versions, compare siblings, merge improvements, and preserve invariants while changing other dimensions. Current AI interfaces make this work unnecessarily hard. As I wrote in Creation as Exploration and Discovery, proper exploration of latent space requires pathcraft, and current AI systems lack adequate UI support for navigating the forkgraph.

Problem-solving in complex domains will usually require branching out among many paths, making a linear scrolling chat a poor UI.

Future AI systems should therefore treat branching as a first-class object. Every meaningful experiment, prompt, candidate, and revision should have provenance. Users should be able to see what changed, why it changed, which result it produced, and how it relates to sibling attempts. The system should automatically group near-duplicate explorations, label major strategy shifts, and let users bookmark not only outputs but also paths. In AlphaEvolve terms, users need to know not just the best program, but the lineage of the idea that produced it.

This applies equally to simpler AI projects. A designer exploring logo concepts needs to branch by style, color, symbolism, and typography. A product manager exploring roadmap strategies needs to branch by market, resource constraint, risk tolerance, and customer segment. A writer exploring article drafts needs to branch by thesis, audience, tone, and evidence. The forkgraph is the natural geometry of AI interaction.

Discovery will usually involve forking the user’s original ideas along multiple branches, including backtracking and the spawning of new subbranches. The UI should support this behavior rather than assuming a linear path through the latent intent space.

However, as AI drives the cost of generating new branches to near-zero, a new UX crisis emerges: the curation bottleneck. If users can spawn a thousand variations in an hour, human attention becomes the scarcest resource in the system. The next frontier of the forkgraph is not just facilitating endless divergence, but building intelligent convergence mechanisms: tools that actively synthesize overlapping branches, highlight thematic clusters, and prune cognitive dead ends, preventing the user from drowning in their own latent space.

We need an improved UI for curating extensively branching explorations.

Look Locks for Science

When writing about creative use of AI, I described the need for “Look Lock”: the ability to freeze certain aspects of an AI creation while exploring other dimensions. In image generation, this might mean preserving a product’s appearance while changing lighting or camera angle. In scientific AI, the same concept becomes constraint lock or metric lock. The user wants to explore new candidate solutions while holding invariant the parts that must not change: validity checks, safety constraints, evaluation boundaries, or domain assumptions.

The AlphaEvolve study shows why this is necessary. Users were not always sure whether they could manually edit certain parts of their experiment, and some found the separation between initial program and evaluation code confusing. Worse, the agent could rewrite code destructively when asked to update one function. These are classic AI UX failures: unclear boundaries, weak preservation of invariants, and insufficient user control over what the AI may modify.

Future AI tools need explicit locks. “You may change the search strategy, but not the evaluation function.” “You may simplify the code, but not weaken the constraint checks.” “You may propose alternative metrics, but do not replace the current metric without approval.” These locks should be visible, manipulable, and persistent across branches. In the AI era, constraints are not dull administrative details. Constraints are the user’s intent made operational.

New Metrics for AI UX

Traditional usability metrics are insufficient for this kind of system. Time-on-task is ambiguous when the AI may run for days. Fewer clicks are irrelevant if the user loses the ability to understand what happened. Satisfaction is incomplete if the user likes the output but cannot trust the process. In Intent by Discovery, I argued that AI UX requires new metrics such as time-to-correct, separation of user slips from system misinterpretations, and perceived agency.

The AlphaEvolve study suggests several additional metrics for discovery-oriented AI. One is time to first viable experiment: how quickly can a user move from vague intuition to a testable setup? Another is time to detect meaningless progress: how quickly can the user see that the AI is gaming the metric or wandering randomly? A third is cost per validated insight: not cost per output, because outputs are cheap, but cost per result that the human trusts. A fourth is branch recoverability: how easily can users return to an earlier promising path without reconstructing it from memory?

We should also measure intent convergence. At the beginning, the user’s intent is vague. After several iterations, it should become sharper. The interface succeeds when the user can say not only “I got a result,” but also “I now understand what I was trying to ask.” This is a profound shift in UX evaluation. The interface is not merely helping users execute intent. It is helping them form intent.

Design Rules for Future AI

The AlphaEvolve paper points to a broader design agenda, as we race to build Agentic AI systems capable of taking over complex, multi-step, autonomous workflows in fields like software engineering, financial modeling, legal research, pharmaceutical discovery, and generative architectural design.

These systems cannot be governed by a prompt box and a scrolling transcript. For serious human-AI collaboration, the interface must be built around the intentmaking–sensemaking loop, mediated by boundary objects and supported by diagnostics, provenance, and controllable constraints. The following design rules generalize from the AlphaEvolve study:

1. Transition from Assistant to “Scientific Instrument”

We must discard the metaphor of the AI as a human-like assistant. If a human assistant returned 30,000 variations of a project in 24 hours, they would be immediately fired for generating catastrophic noise. An AI is an instrument of immense computational scale. UX design for AI should mirror the dashboards of complex physical machinery or scientific instruments. They should offer macroscopic dials, readouts, evolutionary trajectories, and diagnostic monitors, empowering the user to step back and steer the system rather than attempt to converse with it. As the researchers note, using an AI to reduce the cost of trying out ideas is as important as using it to generate ideas.

Advanced AI users need instruments for exploration, not assistants for conversation.

An instrument exposes signals, requires calibration, and improves the expert’s ability to see. The microscope did not replace the biologist but changed what biologists could observe. Similarly, AlphaEvolve changes what mathematical search can reveal, provided that the interface helps the mathematician calibrate, inspect, and interpret the system.

AI should be considered an instrument for exploring the user’s possible goals, not an answer box that provides a single solution (except for simple problems).

This metaphor also avoids the false dream of zero-learning AI. The fantasy is that users will state a wish, and the AI will silently do the right thing. That will work for trivial tasks and is a viable usability goal in those cases. It will fail for meaningful work. Meaningful work requires judgment, and judgment requires visibility. If users never learn how a system behaves, they cannot steer it. In Intent by Discovery, I warned that zero-learning AI risks cognitive offloading and deskilling. Good AI UX should reveal enough structure to preserve human agency.

The DeepMind paper is important because it turns this philosophical claim into observed user behavior. Expert mathematicians did not want a magic answer box. They wanted to play with the system, test ideas, see whether progress was real, discover loopholes, and understand promising branches. One participant compared the experience to playing a computer game and exploring a world. Play and exploration are exactly the right metaphors. The future of AI UX is world exploration.

In complex domain spaces, an expert often learns more from seeing why a path failed than from seeing why a path yielded a temporary high score. To push this calibrated instrument further, systems must also expose a negative solution space. True intentmaking flourishes when an expert can easily audit what the AI rejected. By offering automated, high-level summaries of dead mutation paths, the UI transforms catastrophic programmatic noise into a clear map of boundary constraints, letting users cross-examine the AI’s hidden negative assumptions.

2. Prioritize and Elevate the Experimental Loop

Future AI tools must expect, and seamlessly support, heavy human iteration. Users should never feel penalized for getting a prompt wrong. UX must radically lower the barrier to entry by providing zero-cost sandboxes in the form of rapid testing environments where users can evaluate a tiny slice of the AI’s behavior before committing to a full-scale runs that last days or weeks.

Discovery-oriented work emerges from many small experiments before one large commitment.

For instance, in an AI-driven software architecture tool, users should be able to run a 10-second logic test of an AI’s proposed database schema before allowing the AI to spend an hour coding the backend infrastructure. The UI must encourage a fail-fast mentality for intentmaking, treating the initial AI-generated specification as a disposable starting point.

3. Design Explicit and Intuitive Boundary Objects

Every AI application domain must identify its ideal boundary objects and elevate them to the forefront of the UI. If an AI is generating music for a professional musician, the boundary object is not the underlying MIDI code; it is a visual waveform or interactive sheet music. (For an amateur song-creator like myself, simpler boundary objects would be needed.) If an AI is optimizing a global supply chain, the boundary object is an interactive geographic map, not a spreadsheet of logistical weight-distribution arrays.

Furthermore, AI-generated natural language summaries must be standardized as a connective tissue. Users should be able to hover over any complex machine output and instantly receive an English-language strategic rationale. The UX must dynamically translate the AI’s native language into the user’s native language.

AI systems should support multi-level sensemaking. Users need summaries, visualizations, raw artifacts, provenance, comparisons, and drill-down paths. No single representation is sufficient because users move between overview and detail as their understanding improves.

4. Build Robust Provenance and Versioning Control

One of the major points of friction identified by the mathematicians during the study was the immense cognitive overload of managing dozens of slightly modified experiments. “Telling my experiments apart is really difficult and I have to click into the experiment to remind myself what was different,” one user noted.

As AI allows users to rapidly spawn variations of ideas to dial in their intent, the UX must borrow heavily from Git-style version control. Visual provenance trees that show exactly where an experiment was forked, what parameters were changed, and how the outcomes diverged, will be a mandatory feature. Future AI operating systems must provide a highly visual “multiverse” view of a user’s intentmaking journey to prevent them from becoming lost in their own data. They need a better view, like an overview tree of modifications.

AI systems should replace linear chat history with the forkgraph. Branching, grouping, labeling, restoring, comparing, and locking should be fundamental UI operations. This will matter more with every increase in AI capability because more powerful AI creates more alternatives, and abundance without navigation becomes a swamp.

5. Enhance Diagnostics to Combat Reward Hacking

Agentic AI will inevitably pursue dead ends and engage in reward hacking in any domain, whether it is optimizing tax preparation or engineering a bridge. The UX must provide robust diagnostic tools to help users answer fundamental sensemaking questions at a glance.

To aid this, AI platforms must introduce “Critique Agents” as a standard feature: secondary, adversarial AI models whose sole job is to probe the user’s constraints and the primary AI’s logic to prevent costly hallucinations or logic hacks. Dashboards must include automated cross-check metrics that alert the user if an AI’s performance spike is statistically suspicious. Tools to debug metrics, such as multi-objective evaluation, must be prioritized.

6. Support Conceptual Inquiry Without Destructive Edits

Users need the ability to ask the AI conceptual questions about the ongoing experiment without altering the experiment itself. In the DeepMind study, users complained that asking the AI to update or explain one function sometimes “destroys the rest of the evaluation code.” Furthermore, the conceptual boundaries of the system were not always clear to first-time users. Future UX must clearly separate the inquiry channel (safe brainstorming and sensemaking) from the execution channel (modifying the boundary objects and underlying computational parameters).

The Larger Lesson

The DeepMind study was about mathematical discovery, but its implications are much broader. It shows that advanced AI changes the basic unit of interaction. The user no longer arrives with a fully formed intent and merely asks the system to execute it. Instead, the user arrives with a partial ambition, a hunch, or a problem space. Through repeated interaction with the AI, that ambition becomes sharper.

AI’s near-term contribution to science may be that it dramatically lowers the cost of trying ideas humans already have but cannot afford to test. That is a usability revolution. By reducing the cost of exploration, AI changes which thoughts are worth pursuing. The parallel with ecommerce is simple: when the interaction cost falls, behavior changes.

Lower barriers to use = more use. Basic human nature, whether we’re talking about mathematicians in the top 0.1% of the IQ curve or regular folks shopping for a new outfit.

This also explains why a math paper aligns so tightly with my work on creation by exploration and discovery. In that article, I argued that AI moves creation from building to describing and finally to discovering. Users will navigate latent solution spaces rather than specifying outcomes in advance, and the main UX problem will become helping people recognize, compare, and refine possibilities. AlphaEvolve shows the same transition in a rigorous domain. The mathematicians were not merely describing the desired solution. They were discovering both the solution and the proper formulation of the search.

The future of AI UX, therefore, is not better prompt boxes. Better prompting is only the bridge from the current era to the real one. The real interface is an exploration environment: a structured, navigable, branching space where the user’s intent becomes clearer through interaction. DeepMind calls this intentmaking. I have called it intent by discovery. The names converge because the phenomenon is real.

AI systems that understand this will win. They will not ask users to produce perfect prose from a blank screen. They will give users alternatives to react to, constraints to lock, branches to compare, progress to inspect, and loopholes to close. They will support the human not as a typist of prompts, but as an explorer of possibilities.

To unlock AI’s immense potential across all industries, from medicine to engineering to the creative arts, AI-UX designers must embrace the messy human reality of problem-solving. By abandoning conversational interfaces and instead building rich, dashboard-driven environments centered entirely on the Intentmaking and Sensemaking loop and anchored firmly by intuitive, domain-specific Boundary Objects, we can transform AI from an unpredictable digital assistant into a robust, world-altering instrument. The future of discovery is not fully automated; it is profoundly, iteratively, and beautifully collaborative.

The DeepMind study was about scientific intent discovery, but the same conclusion applies to any AI used in complex domains by advanced knowledge workers.

The most important output of advanced AI may not be the answer. It may be the user’s improved understanding of the question. That is the deeper meaning of intentmaking. In advanced AI, intent is not just an input to the interface. It is one of the interface’s most valuable outputs.

Intentmaking (shape what we’re doing) and sensemaking (understand the outcome): the two sides of AI for advanced knowledge workers.

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 (30,428 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 June 1, 2026 (GPT-Images-2)

Ecommerce Usability Fails

You would think that after 30 years of building ecommerce sites, we would know how to design an ecommerce website for maximal usability and conversion. You would be right — and wrong!

(I refuse to abandon the lovely em-dash just because most AI models overuse this useful typography element.)

We do know the usability guidelines for ecommerce design, but they are often overlooked, especially by smaller, specialized vertical sites that could benefit the most from good design.

Two of the most egregious, sales-killing usability mistakes in ecommerce are found in the checkout process, which is otherwise the best-documented part of ecommerce design: Coupon codes and required registration without a guest checkout option.

A highly prominent coupon code field directly on the checkout page is often included to please VIP shoppers or run seasonal promotions. But coupon fields act as a severe conversion barrier for the vast majority of regular visitors. By prominently displaying an empty discount box, you are actively interrupting a seamless purchasing flow and introducing unnecessary psychological friction.

When a prepared user without a promo code sees this field, they immediately experience the fear of missing out. The implicit message is that the current price is a bad deal, prompting the shopper to pause their transaction. Consequently, they quickly abandon your checkout page and leave the site to search for coupons on the Internet.

This external search represents a massive risk to your bottom line. Search engine results for brand discounts are notoriously treacherous. Your prospective customers will inevitably encounter affiliate marketing traps, frustrating expired codes, and most dangerously, highly targeted advertisements from your direct competitors. This distraction often leads them to find another site where they buy a very similar product at a cheaper rate.

Even if shoppers resist the lure of a competitor and return to your website, the damage is already done. Returning without a working code, they must face reality. They feel cheated because they have to pay full price while others get a discount. Ultimately, this profound sense of unfairness may lead them to not buy at all.

Showing a coupon code field on an ecommerce checkout page is like opening a portal to the rest of the Internet, encouraging the prospective customer to step through in order to find a deal, which they’ll often close on a different website. (GPT-Images-2)

The solution? Automatically add the discount to the cart when users click through from a promotion. You save the loyal customer the extra step of copy-pasting the coupon code and gain the sales benefit of reminding users about the discount on every single product page where the promotion applies.

It’s just as bad when an ecommerce site requires account creation from new shoppers instead of allowing guest checkout. An online shopping experience should be entirely effortless, but forced signups introduce immediate, frustrating barriers.

The mandatory registration process is usually convoluted, including demands for a unique userid where all variations of the user’s name are usually taken by older (maybe even inactive) accounts. This effectively forces new customers to invent confusing, unnatural usernames they will never remember.

The next painful step is password creation with so many special requirements that users repeatedly fail to meet the platform’s overly strict security rules (password pain was number 6 on my list of top-10 UI annoyances). Discovering the exact required combination of uppercase letters, obscure symbols, and numbers is mentally exhausting. Extremely frustrated by these constant, unnecessary digital roadblocks, often people give up and abandon the registration process entirely. They leave their full shopping carts behind, permanently costing the site the sale.

A second major conversion killer: the lack of guest checkout, imposing the pain of convoluted account creation. Users often give up once they get slapped down too many times. (GPT-Images-2)

Fortunately, the optimal solution is straightforward. Instead, offer smooth guest checkout without the friction of account creation. Let the buyer pay without jumping through administrative hoops. Then after the user has completed the purchase (and the site has closed the sale and made the money), offer an easy way to convert the data the user has already entered into a registered account for easier subsequent purchases. This smart approach prioritizes seamless customer experiences, significantly reduces cart abandonment, and ultimately guarantees your ecommerce store’s lasting, highly profitable online business success.

Repeatedly encountering these two design mistakes even today offends my usability sensitivities because I have explained them so often over the decades. But there are many more: The Baymard Institute has documented more than 700 of UX guidelines for ecommerce design. The full subscription costs $400 per month, which expensive, but should pay for itself for any site with more than minimal sales.

However, if you are cheap, of if management doesn’t believe in user research, Baymard now also has a free option, where they publish abbreviated guides to their most important research findings. (I think this is smart for their sales, because after a company redesigns to fix a few usability flaws, the sales increase will often be sufficient to motivate even the most usability-hostile leadership to cough up the funds for more insights.)

Check out the 4,000 word article about Baymard’s checkout usability findings. This is about as long as my articles, so it’s worth an entire infographic series to summarize:

(GPT-Images-2)

Social Media Dominates Media Consumption, But Is Not Social

Market research firm GWI has published an interesting analysis of many of its recent surveys called Connecting the Dots 2026. It’s surveys on consumers’ media consumption places social media on top, closely followed by short videos (e.g., TikTok), for a total of 13.5 hours spent on average with these two media forms. In contrast, traditional broadcast TV clocks in slightly below 5 hours per week. (This across 54 countries where GWI conducts market research.)

Weekly media consumption, according to GWI’s consumer survey. (GPT-Images-2)

Slightly scary: young people (16–24 years old) spend even more time on social media and short videos: 18 hours weekly. What a way to fritter away your youth, confirming the old saying that youth is wasted on the young.

However, despite the traditional name of “social” media, social media use has become less social in recent years, according to GWI data. Ten years ago, the top three reasons to use social media were to stay in touch with family & friends, to meet new people, and to share the user’s opinions with others. All strongly social motivations for using the platforms.

Now, the top reasons are to fill spare time or relax, to find entertaining content, or to see what’s trending, none of which relate to connecting with other people or being social.

Is this good or bad? Probably mainly bad, contributing to feelings of being lonely and isolated. A redeeming point, also from GWI research, is that AI seems to be breaking this trend toward passive and dulling media consumption to make users more active and creative.

Agent Oversight Usability

The history of user interface design is a story of shifting the burden of labor from the human to the machine. We evolved from punch cards to command lines, and from graphical interfaces to touchscreens, always seeking to reduce friction. Yet, in the nascent days of generative AI, we took a frustrating step backward. We suffered heavily through the “blank page” problem. Users were forced to painstakingly articulate every micro-step to a chatbot, desperately trying to engineer the perfect prompt to coax out the desired result. The human was doing the heavy lifting, acting as a translator for a machine that lacked initiative.

Steadily improving user interfaces, from text-based command lines to graphical user interfaces, have shifted much of the burden from the human to the computer. AI agents are the next step in this sequence of load-shifting. (GPT-Images-2)

No more. We are entering the era of autonomous AI agents: systems designed to take a high-level goal and execute complex, multi-step workflows without the need for constant, explicit prompting.

For UX professionals, this means the primary UI for AI is no longer conversational. Users are now managing ecosystems of agents. Just as designers were getting a handle on how to increase the usability of the ubiquitous chat window (erroneously heralded as the ultimate, frictionless interface), it is rapidly becoming a legacy bottleneck. Today’s users are no longer simply chatting with a singular, subservient bot; they are deploying, monitoring, and managing sprawling ecosystems of specialized agents working in parallel.

Conversational user interfaces were never a guarantee of usability. They were still hard work. (GPT-Images-2)

The new usability challenge: make it easy to oversee many AI agents, each running autonomously, and understand what each is doing. (GPT-Images-2)

Consequently, the fundamental usability challenge is inverting once again. We have shifted away from the problem of articulation (how do I meticulously say exactly what I want the machine to do?) to the far more critical problem of oversight (is this autonomous entity actually doing what I intended, and is it doing it safely?). Consider the stakes in this new reality: an agent can now autonomously refactor an entire monolithic codebase, independently research and negotiate procurement contracts, or execute a multi-channel, dynamically optimized marketing campaign while the user is away from the keyboard. When software operates with this unprecedented degree of agency, traditional step-by-step interaction models fundamentally fail.

The stakes for getting AI usability right keep getting higher as AI agents tackle bigger problems and run for days or weeks. (GPT-Images-2)

From the articulation barrier to the oversight burden: usability in the AI era keeps shifting under our feet. (GPT-Images-2)

In this new paradigm, the user’s role is elevated from a micro-manager to a Creative Director. A Creative Director doesn’t push every pixel or write every line of copy; they set the vision, review the progress, and provide critical course correction. But a director cannot lead if they are locked out of the studio. If our new interfaces fail to provide high-level state visibility, perhaps the most vital of all core usability heuristics, the user will inevitably lose trust in the system. In the realm of autonomous AI, trust is the entire product.

Therefore, our mandate as designers must evolve. We must move beyond the simplicity of “Send” buttons and linear chat histories. We must design interruption points and explicit review gates. We need to engineer control towers that allow humans to pause, inspect, and course-correct an agent mid-flight. If the user cannot easily see, interpret, and logically follow the agent’s “thinking” at scale, they are flying completely blind. And if they are flying blind, they are no longer in charge of the user experience; the machine is. It is time we design interfaces that restore the user’s control, allowing them to confidently govern the machine rather than just talk to it.

My main lifetime goal has been to place humans in charge of computers instead of being subservient. AI is no different. But what’s new is that AI now enables us to run multiple autonomous agents, meaning that we must change from an operational user interface to an oversight UI. (GPT-Images-2)

Your Team as Video Game Characters

Idea for team-building or to introduce colleagues in other disciplines to your team: draw them as characters in popular video games.

Here, “what if UX Designer were an NPC in GTA6?” (for extra credit and maybe too-much realism, you can upload the person’s likeness as reference images). Repeat for different roles and sub-disciplines.

UX designers as non-player characters (NPC) in Grand Theft Auto 6. (GPT-Images-2)

Video game metaphors don’t have to be only about people or work roles. We can also visualize concepts or theories that way, though my attempts with GTA screens have been less successful. Here are some that worked:

Usability Heuristic 7 (Flexibility and Efficiency of Use) as a GTA character. (GPT-Images-2)

Progress indicators. (GPT-Images-2)

What’s a Good Rating Scale Score?

I recommend against heavy use of rating scales in user research, because they only measure what users say about usability, not what they actually do, and thus don’t capture true usability or task performance. Much better to spend most of your research dollars observing users’ behavior, whether you watch personally or have an AI watch for you.

That said, there are also benefits to knowing how much users like your design, and so I don’t want to abolish rating scales completely. But when you get back a score on a 1–7 scale, how do you know whether that number is good or bad, and what to do about it?

The best way to know is to use a consistent survey instrument repeatedly over the years, so that you can track the scores. If the score is up from last time, then people liked your redesign! Simple.

The simplest way to interpret survey results is to compare them with the outcome of exactly the same survey asked about your previous design iteration. (GTP-Images-2)

Second-best: use a standard question that other teams have used many times before to capture user opinions about other designs. If you score in the top 25% of those previously-collected measures, you have a design that’s pretty good, but if you score in the bottom half, you’re clearly worse than most other designs. I recommend using the Single Ease Question (SEQ) from MeasuringU: substantial benchmark data has already been collected for this one-question survey, and because it is indeed limited to a single question, administering it doesn’t impose a major burden on your users.

When you use a standard survey instrument, you can compare your results with those in the benchmark. (GPT-Images-2)

Surveys should be as short as possible. Every extra question lowers your response rate and thus undermines the validity of the answers. (GPT-Images-1)

But what if you need unique insights that can’t be captured by a standardized survey developed by outsiders to satisfy a wide range of design projects? You have a new question that you never asked before, and which nobody else has asked either. No old data for comparing your newly collected scores!

Luckily, my old friends, Drs. Jim Lewis and Jeff Sauro from MeasuringU provide a rule of thumb for judging rating scales for which no historical data is available.

Very simple:

• Good score: top 20% of the scale, so for a 100-point scale, 80 or above, and for the traditional 1–7 scale, 5.8 or above.

• Average score: around 70% of the range of the scale, so 5.2 on a 1–7 scale.

• Poor score: below the midpoint of the scale, so 4.0 or less on a 1–7 scale.

The rule of thumb for judging rating scale scores, here converted to the common 1–7 scale used in most user research surveys. Notice how skewed it is, relative to the mathematical mean of the scale. (GPT-Images-2)

Why is the mathematical midpoint of the scale poor rather than average? Because users tend to be polite and score generously. Acquiescence bias (also called agreement bias or “yes-saying”) is extremely well documented, also outside usability research. Humans have a systematic tendency to agree with survey statements or answer “yes.”

Humans aren’t math. Interpret survey scores accordingly. (GPT-Images-2)

A classic indicator is when respondents tend to endorse both sides of logically opposed statements, such as “I like taking charge” and “I prefer to let others lead,” at rates higher than would be expected if they were answering carefully.

Acquiescence bias makes survey responses hard to interpret correctly. (GPT-Images-2)

Acquiescence bias has several underlying drivers: cognitive ease or effort reduction (agreeing is a low-friction default), social desirability (wanting to appear polite, cooperative, or non-confrontational), and cultural norms that reward deference or agreeableness, especially in Asia. Sociological studies also emphasize status differentials and perceptions of the researcher as a higher-authority figure, which can lead particularly lower-status respondents to “yea-say” out of respect or the wish to avoid conflict.

A mathematically average survey score is thus a bad score in terms of the underlying experience factors you’re trying to assess.

If you must use subjective rating data, at least follow the guidelines in this comic strip. And combine with observational data. (GPT-Images-2)

Heroes of UI: Ben Shneiderman

Ben Shneiderman is a pioneer of user interface research and one of my heroes.

Ben Shneiderman career overview. (GPT-Images-2)

The thinking image model shines in pulling together information from this kind of infographic. I am disappointed that it didn’t show the real covers of Shneiderman’s books. They can absolutely be found on the web. Personally, I would also have featured his early book Software Psychology (1980) which was a pioneering work that influenced me when I started. As a book author myself, I’m sure that Ben prefers AI to highlight his latest book (featuring his latest thinking) instead of one of his first (featuring outdated information), but to be true to history, Software Psychology was the more important work.

Seedance 2.0 Concept Advertisement

What if the leading AI video model had been a packaged software product in 1980? Here’s a concept ad that could have run in the personal computer magazines of the era:

(GPT-Images-2)

Considering the expense of the AI credits to make videos with Seedance (it’s by far the most expensive video model, but it is also the best), I would have saved a lot of money if I could have bought it as packaged software at the advertised price.

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 (30,404 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)

(All images in this article made with GPT-Images-2)

The OpenAI Deployment Company (often called DeployCo) is OpenAI’s new consulting venture to help enterprises design, build, and productionize AI systems using OpenAI’s AI models. They recently announced a $4 billion investment in DeployCo and partner agreements with several big (if legacy) consulting companies, including Bain & Company, Capgemini, and McKinsey & Company. More interesting than the old-school collaborators is the deal for DeployCo to acquihire 150 Forward Deployed Engineers from the more leading-edge consulting company Tomoro.

A Forward Deployed Engineer (usually abbreviated “FDE”) is a software engineer who does not sit in a siloed product team at a remote headquarters. Instead, the FDE embeds directly with client organizations to design, customize, and deploy a company’s technical platform in the client’s real operating environment. This is one of the best ideas to come out of Silicon Valley recently, because the FDE becomes a discovery mechanism for features to build into the product.

Sitting with customers on the front lines and experiencing firsthand what works and what fails beats any amount of speculation by designers or product managers sequestered back at headquarters. This is especially true for domain-specific, complex enterprise apps, whether that involves tracking logistics on a massive shipping port, managing supply chain reconciliation in a pharmaceutical plant, or navigating compliance auditing in a global financial institution. Context is king, and the FDE breathes the context daily.

Context is King: being on-site beats sitting back at HQ and speculating on what customers need.

But context has a dangerous side. The person who lives inside a workflow can become over-adapted to it, accepting the organization’s weird rituals as facts of nature rather than fossils of past constraints. The best field work does not merely absorb context; it must also create enough distance to challenge context. This is the crucial difference between being an embedded outside expert and being existing company staff.

Question the rituals. Tradition is rarely a reason to keep doing things the old way, especially when conditions change radically, as they do with AI.

DeployCo claims that their FDEs will have access to the next generation of AI models from OpenAI, enabling them to build customer solutions that leverage where AI is going rather than where it has been, as most enterprise IT teams do.

The Trap of Local Optimization

Embedding technical talent directly with clients is a brilliant model for traditional software, but engineering alone isn’t enough to unlock the true value of generative AI. In traditional software design, we aim to identify users’ pain points in their existing workflows and design a UI that enables them to get their work done faster and more easily. If an accounts payable clerk spends three hours manually matching invoices to purchase orders across two clunky systems, traditional software attempts to digitize, integrate, and streamline that manual process.

A good FDE can spot these friction points in the field and communicate them to a UX team back at HQ, where they can be transformed into polished product designs. (Some of that polish is added through traditional user testing, where we easily see if, for example, a command name violates the users’ mental model. If so, change it and save a few hours of training time.)

With AI, the old workflows must no longer be treated as the design brief; they must be questioned at the root. AI is, in fact, a great productivity enhancer even when used in the traditional way to increase the efficiency of individual employees performing the same tasks as always, just faster and better. A paralegal can summarize a legal brief in seconds; a junior developer can write boilerplate code instantly; a digital marketer can generate campaign copy with a single prompt. We can typically improve that employee’s performance on those specific rote tasks by roughly 40%.

But at the company-wide level, such local productivity gains rarely translate into substantial profit gains and shareholder value. When you have a long chain of steps and optimize only a few, the delay simply shifts to the remaining steps, which will dominate the overall time to solve the underlying problem.

Local fixes to specific tasks are fine, as far as they go, but they don’t go far enough to revolutionize business. We must take a wider view and redesign the full workflows for the new AI opportunities.

Consider an enterprise commercial loan approval process. The workflow consists of data gathering from the client (Task A), initial risk assessment (Task B), legal compliance review (Task C), pricing formulation (Task D), and final managerial sign-off (Task E).

If an FDE embeds with the initial risk assessors and builds an AI-powered tool that speeds up Task B by 40%, the business may assume it has achieved a corresponding operational win. But if Task B was never the ultimate systemic bottleneck, the overall cycle time won’t improve. If Task C (the legal compliance review) remains constrained by traditional, unoptimized human limits, the newly accelerated risk assessors are simply piling up completed assessments at the door of the legal department faster than ever before. The overall time it takes for the customer to receive a loan decision may only have been shortened by a few minutes. You have fundamentally changed nothing about the business model; you have merely hurried up the process of waiting. The business value of that 40% speed boost on Task B is swallowed by the friction of the legacy workflow.

Increasing the speed of one step won’t buy you much overall efficiency if the other steps remain slow.

The correct unit of analysis is therefore not task efficiency but system throughput. A company does not get paid more because one employee finished one intermediate activity faster. Profits only grow when the entire value stream moves faster, with fewer defects, less rework, and less managerial expediting. Local AI often increases work-in-progress inventory: more partially completed cases, more pending approvals, more items waiting in queues. This can even make the organization feel busier while making the customer experience no better. The FDD’s job is to redesign the flow, not celebrate the acceleration of a single eddy in the stream.

AI-Native Workflow Redesign

Full workflow redesign is required to take advantage of AI’s new capabilities for removing steps entirely and designing human limitations out of the system.

Historically, corporate workflows were designed to accommodate human cognitive and physical limits. (In fact, that was our traditional goal in human factors engineering — thus the name.) Humans are mostly serial processors: we do one cognitively demanding thing at a time, have sharply limited working memory, require sleep, and operate within work schedules, burning out if we work more than 60 hours per week for any length of time. Well-designed legacy corporate processes reflect these constraints.

We used to design to overcome human limitations, which have remained constant over the decades, leading to stable usability guidelines. Now, we will be designing harnesses to compensate for AI limitations, which change annually, if not faster, so our UX guidelines for AI design will likely need to be updated frequently.

We introduced rigid approval hierarchies because a single human can only review so many documents in an hour and is likely to make errors. We created siloed departments because human expertise is narrow and specialized. We built complex status dashboards because human managers cannot keep the real-time state of 10,000 concurrent operations in their heads.

Agentic AI systems do not share these limitations. An advanced multimodal AI model can read a 500-page regulatory filing in seconds, cross-reference it against 10 years of historical company data, verify it against complex policy documents, and make an instant probabilistic risk assessment. It can process thousands of these files in parallel. And it works 24/7, vastly facilitating international business operations across time zones.

If we simply use AI to help a human read the document faster, we are leaving the vast majority of the technology’s potential on the table. We must completely abandon the old workflow. The new workflow shifts the human out of the data processing phase altogether, elevating them to the role of an orchestrator, strategic decision-maker, or exception-handler.

Once AI removes the cognitive bottleneck, a different bottleneck appears: authority. The limiting question becomes not “Can the system know what to do?” but “Is the system allowed to do it?” AI-native workflow design must therefore redesign decision rights, escalation rules, audit trails, and accountability boundaries. Otherwise, the organization merely replaces slow human cognition with fast machine recommendations waiting for the same old human permission structure.

AI is fast. Humans are slow. This means that if humans have to approve every last thing the AI has done, we won’t realize the needed workflow efficiencies.

Permission structures will need continued redesign as AI becomes more capable with each release. As the banner says, it’s a design choice which AI actions require human review. Too many, and nothing gets done, plus safety is undermined by overwhelming the humans, leading them to approve without sufficient reflection.

I don’t believe most FDEs are capable of such revolutionary workflow redesign. Their training, incentives, and core competencies are rooted in technical problem-solving, systems architecture, algorithmic optimization, and code execution. Give them a defined problem, and they will build a flawless technical solution. But asking them to fundamentally reimagine the operational blueprint and human dynamics of a Fortune 500 company is outside their mandate. They are trained to optimize the “how,” not fundamentally question the “why.”

The obvious objection is that the best FDEs will say they already redesign workflows, and in a limited sense they do. There’s no monopoly on human factors thinking. But engineers’ redesign typically begins from the computational affordances of the platform: where data resides, where APIs connect, where latency appears, where a model can be inserted. The FDD begins one level higher, with the social contract of the work: who is allowed to decide, who must be informed, who bears risk, and which organizational ritual exists only because yesterday’s tools were dumb. That is not a smaller version of engineering. It is a different starting point.

Equally, I don’t believe most current UX professionals can do this either, limited as they are in their thinking by years of designing pre-AI solutions. For four decades, the UX industry has been obsessively focused on screen design, click-reduction, visual hierarchy, and graphical user interfaces (GUIs). Traditional UX asks, “How do we make this form easier for the human to fill out?” AI-native design asks, “Why does this form exist at all, and can an autonomous agent negotiate this data exchange in the background without human intervention?”

However, workflow redesign at this scale is, at its heart, a massive service design problem. It requires taking a step back, examining the ultimate goal of the business ecosystem, and architecting an entirely new way to achieve that goal. It should be attacked with solid user research and innovative service-design methods. These are firmly within the broader UX wheelhouse, and I firmly believe the best, most adaptable UX professionals of today will evolve into the AI workflow designers of tomorrow.

Introducing the Forward Deployed Designer (FDD)

To realize the true transformative value of enterprise AI and bridge the massive gap between technical capability and organizational transformation, I suggest inventing a new consulting category. Whether formalized within DeployCo, adopted by legacy consulting companies like McKinsey and Bain, or championed by the inevitable wave of new agile boutique firms founded specifically to exploit AI opportunities: We need Forward Deployed Designers, or FDD.

FDEs are great, but we also need FDDs to realize the full potential of AI for improving company performance.

An FDD is a hybrid professional: part ethnographic researcher, part macro-systems service designer, and part AI product strategist. Like the FDE, the FDD does not sit in an ivory tower at headquarters. They embed directly on the factory floor, in the trading room, in the hospital ward, or within the back-office cubicle farm. But where the FDE embeds to understand the data architecture and API endpoints, the FDD embeds to understand the business intent, the human psychology, and the deeply ingrained legacy constraints of the organization.

What exactly makes a Forward Deployed Designer? It is not simply a traditional UX designer given a trendy new title, nor are they merely an AI prompt engineer. They combine ethnographic observation, systems reasoning, organizational judgment, and enough technical intuition to know which fantasies can be built.

The missing fourth ingredient is political literacy. Workflows are not merely sequences of tasks; they are settlements among departments, professions, budgets, regulators, unions, managers, and status hierarchies. A field designer who cannot read the informal power map will mistake an org chart for an operating system. The real workflow is often hidden in who can say no, who can override the system, who gets blamed when exceptions occur, and which spreadsheet everyone trusts more than the official dashboard.

Map the internal company politics and power structures, which often deviate from the official org chart.

Core FDD Attitudes and Mindsets

• Constructive Irreverence (Radical Deconstructionism): FDDs must possess an almost ruthless willingness to kill existing processes. They cannot be intimidated by “the way we’ve always done it.” They must have the fearless willingness to look a client executive in the eye and say, “We aren’t going to build an AI tool to help your team process these forms 40% faster. We are going to build a system that eliminates the need for these forms to exist entirely.”

• Replace User Advocacy with Outcome Advocacy: Traditional UX rightly minimizes user effort within the current task. The FDD goes one level higher. Sometimes the right design is not to make the current job easier, but to eliminate the task and move the human into governance, judgment, or relationship work.

It’s probably a bitter pill to swallow for many old-school UX folks, but user advocacy is no longer our job. We should not advocate for the outcome of AI processes, many of which will involve no human users.

• Extreme Comfort with Ambiguity: Pre-AI software is deterministic: if a user clicks a button, the exact same thing happens every single time. AI is probabilistic, generative, and non-deterministic. Responses vary based on context; hallucinations happen, even if less frequently, as models improve. The FDD must thrive in this ambiguity, designing systems that accommodate flexible, conversational, or agentic interactions rather than rigid, linear pathways.

Core FDD Skills

• Macro-Systems Thinking: The ability to zoom out from a single user interface and hold a mental model of an entire enterprise’s operational architecture. An FDD must understand how changing a workflow in the procurement department will impact the data pipelines and human cognitive load in the legal department downstream.

• Deep AI Capability Literacy: While they do not need to write production-ready Python code or train neural networks, an FDD must fluently understand the “grain” of AI. They must know the practical limits of context windows, the nuances of harnesses, the realities of data latency, and the boundaries of model reasoning. Without this literacy, they will design science fiction that cannot be built.

• Workflow Observability: FDDs must make invisible work measurable before they redesign it. They should instrument cycle time, queue length, rework, exception frequency, approval latency, handoff failure, data re-entry, and the number of times people leave the official system to get the real job done. Without this observability layer, redesign becomes workshop theater: persuasive diagrams unsupported by operational evidence.

• Rapid Conceptual Prototyping: FDDs must move at the speed of thought. Using off-the-shelf AI, API wrappers, and no-code connection tools, the FDD must be capable of stringing together low-fidelity mock-ups of AI workflows to test human reactions and system logic in days, rather than waiting months for formal engineering.

Uncovering the Invisible: Methods for Observing Users in the AI Era

Before an FDD can redesign a workflow, they must deeply understand the reality of the current state. However, traditional UX research methods, like my beloved lab-based usability testing, where you watch a user click through a prototype and ask what they think, are woefully inadequate for AI redesign. Human beings are notoriously bad at accurately describing their own workflows. They omit steps they perform automatically; they fail to mention ad-hoc workarounds; and they confidently state they follow standard operating procedures when, in reality, they rely heavily on tacit intuition.

FDDs must act as workplace anthropologists, deploying immersive research techniques to uncover the “invisible work” that AI is perfectly suited to absorb. (See the comic strip I made about why UX Anthropologists have the most valuable skills in tech these days.) Tacit knowledge elicitation becomes the decisive research method.

Just as important is decision-rights mapping. The FDD must learn not only what people do, but what they are authorized to decide, what evidence they consider sufficient, when they seek cover from a superior, and which approvals are genuine safeguards rather than inherited ceremony. Many workflows look like information-processing systems, but are actually liability-distribution systems. Automating the information flow while leaving the liability architecture untouched produces timid AI: intelligent enough to recommend, but never empowered enough to transform.

Contextual Inquiry and Cognitive Task Analysis (CTA): Traditional shadowing often focuses on what a user does on their screen (Click A, Open B, Type C). Because AI is fundamentally a cognitive automation technology, the FDD must utilize Cognitive Task Analysis to discover what the user is thinking.

When an experienced insurance underwriter evaluates a complex policy, they aren’t just reading a checklist; they are applying years of intuition, recognizing subtle patterns in risk, and cross-referencing disparate data sources. The FDD’s job is to deconstruct this intuition. What memories are they accessing? What data points are they synthesizing? What specific edge cases cause them to pause? By mapping the cognitive workflow rather than just the physical workflow, FDDs translate human heuristics into system prompts and agentic reasoning frameworks, designing AI agents that augment high-level decision-making rather than just automating basic data entry.

Hunting for “Glue Work” and Shadow IT Archaeology: One of the most fertile grounds for AI workflow innovation lies in a company’s “Shadow IT,” the unofficial, unapproved systems employees invent to get around the limitations of their official software. FDDs aggressively hunt for sticky notes on monitors, sprawling, undocumented Excel macros passed around via email, and informal Slack channels where employees ask each other policy questions.

In legacy workflows, humans often act as the APIs between disjointed software systems. The FDD watches for the “swivel-chair integration moments” where a user exports data from System A, reformats it, and pastes it into System B. This human glue is cognitive drudgery. Any such workaround is an indicator of system failure and a blueprint for where a multimodal AI agent could simply read the screen of System A (or, later, retrieve it through a “headless” API), interpret the unstructured data, and autonomously populate System B.

In “glue work,” the user acts as a human computer and laboriously executes steps to connect different computer systems. AI should do this.

But shadow IT is more than evidence of broken integration. It is often the company’s real ontology. The unofficial spreadsheet reveals the categories employees actually use. The Slack workaround reveals the questions the official knowledge base failed to answer. The sticky note reveals the risk signal no database field was designed to capture. Before replacing these artifacts, the FDD must mine them for meaning. Otherwise, the new AI system will faithfully automate the official process while missing the practical intelligence that kept the business running.

Analyzing the Exception Landscape: In most enterprise workflows, standard operating procedures cover the routine 80% of the work. Humans spend a disproportionate amount of time and mental energy navigating the 20% of exceptions, edge cases, and anomalies. Traditional automation breaks down completely in the face of exceptions. Generative AI, however, thrives on unstructured ambiguity. FDDs should observe how experts investigate anomalies. By capturing the intuition and external research humans use to solve edge cases, the FDD can design AI guardrails that handle a vast majority of future anomalies autonomously.

Outcome Mapping via the “Magic Wand” Technique: Because users are anchored to the constraints of their current tools, if you ask them what they want, they will ask for a faster version of their current tool (the proverbial “faster horse”). FDDs bypass this by using the Magic Wand technique during contextual interviews.

They ask stakeholders: “If you had an infinitely capable intern who knew everything about this company, worked at the speed of light, and never made a mistake, how would you instruct them to achieve this outcome for you? What high-level work would you focus on while they did it?”

Try the magic wand technique in your next user interview.

This shifts the user’s mindset from executing tasks to managing outcomes. It decouples the user’s intent (the ultimate business goal) from their action (the steps they currently take to get there), revealing the true parameters of a successful job and providing the blueprint for what the AI needs to achieve autonomously.

Architecting the Future: Service Design Methods for AI Workflows

Once the existing realities of the work are understood, the FDD does not look for ways to incrementally speed them up. Instead, they pivot to advanced Service Design methodologies to synthesize their findings and build entirely new paradigms that take full advantage of AI capabilities. Service design is the craft of organizing people, infrastructure, communication, and material components of a service to improve its systemic flow. In the AI era, the infrastructure now includes autonomous, reasoning agents.

Zero-Based Workflow Design (First-Principles Ideation): Borrowing from the financial concept of zero-based budgeting, FDDs employ Zero-Based Workflow Design. Instead of taking an existing 15-step process and asking, “Which of these steps can AI do faster?”, the FDD starts with a completely blank slate.

Instead of taking an overly complicated legacy workflow and designing the most annoying parts out, success is more likely to come from a blank-slate approach: start with nothing and add only the most necessary parts.

They define the ultimate business goal, for example, issuing an accurate refund to a frustrated customer within 60 seconds. Then, they ask: “Given today’s frontier AI capabilities, if we assume an AI can instantly read the policy, verify the purchase history via an API, and generate the necessary database queries, what is the absolute minimum number of steps required?” The answer could easily be a reduction from 15 steps to two. Zero-based design forces the organization to let go of its attachment to legacy processes and prevents it from paving the cow path (automating inherited bureaucracy). It recognizes that using an AI to write an email, which is then sent to another human who uses AI to summarize that email, is an absurd waste of energy. The new workflow dictates that the two AI systems simply negotiate the refund directly via API.

This is a paradigm change for UX. The interface is no longer always a screen shown to a person. Sometimes the interface is a contract between agents: the data schema, confidence threshold, policy constraint, audit log, exception trigger, and permission boundary that allow two systems to complete work without dragging humans through the middle. In AI-native enterprises, designers will increasingly design service encounters that no customer, employee, or manager ever directly sees.

AI-Augmented Service Blueprinting: The core architectural tool of the FDD is the Service Blueprint. A traditional service blueprint maps the customer journey across the top, showing the “Front Stage” (what the customer sees) and the “Back Stage” (what employees do behind the scenes to support the customer).

The FDD introduces a radical update: the AI-augmented service blueprint. This framework introduces new, critical lanes:

• Front-Stage AI (The Copilot): AI systems that interact directly with the customer or collaboratively with the front-stage human in real-time.

• Back-Stage AI (The Autonomous Agent): AI systems running asynchronously in the background, executing long-chain tasks, monitoring unstructured data streams, and preparing environments before the human even knows they are needed.

Front-stage AI is AI that interacts with a user. But most of AI design will be back-stage AI that works on its own, without humans.

By mapping these lanes together, the FDD visually demonstrates how the workflow shifts from “human doing the work, assisted by a machine” to “machine doing the work, governed by a human.” It clearly defines where entire columns of traditional back-stage human processing can be eliminated, and where human-AI handoffs must occur.

Capability Matching and Role Re-architecture: In a pre-AI world, human job roles were dictated by the limitations of technology. A “Data Entry Clerk” job only exists because computers couldn’t read unstructured handwriting.

The FDD uses a matrix called Capability Matching to challenge this. They map the required steps of the newly designed, zero-based workflow against the unique strengths of Frontier AI (pattern recognition, infinite patience, instant synthesis, multilingual translation, parallel processing) versus the strengths of humans (empathy, moral judgment, relationship building, physical-world presence, legal accountability).

One of the many strengths of AI is its ability to do many things in parallel. Humans can do only one thing at a time, and when they try to multitask, errors explode, and inefficiencies abound.

Tasks are assigned strictly based on this matrix. If a task requires empathy and accountability (e.g., calling a client to inform them their loan was denied), it is assigned to a human. If the task requires synthesizing 400 pages of financial history to determine if the loan should be denied, it is assigned to the AI. This process inevitably results in the complete rewriting of HR job descriptions as a natural byproduct of the design process.

Designing the “Human–AI Handoff” (Friction as a Feature): Because generative AI models are probabilistic, meaning they can confidently hallucinate or make incorrect moral judgments, workflows cannot be fully autonomous out of the gate, especially in enterprise environments burdened with compliance and risk. Therefore, one of the most critical elements of AI service design is the point at which the AI stops, and the human takes over.

Traditional software design strives for seamlessness. In AI systems, FDDs must deliberately practice “Seamful Design.” They design moments of cognitive friction with escalation paths and checkpoints that require the AI to transparently communicate its confidence level and explain its reasoning to the human operator.

We need to retain some “seams” in the workflow to pause and check.

The FDD designs the UI of this handoff not to be a simple “Approve/Deny” button (which inevitably leads to dangerous automation bias, where tired humans blindly approve everything). Instead, it is designed as an evaluative “Proof of Work” interface. The AI must present its conclusion alongside cited sources, identified risks, and alternative options, forcing the human into a critical-thinking mindset before proceeding.

The strongest handoff also shows what would change the answer. A useful AI review screen should not merely say, “Here is my recommendation.” It should say, “Here is the evidence that drove my recommendation, here is the weakest part of my reasoning, here is the missing information that would reduce uncertainty, and here is the counterfactual evidence that would make me reverse course.” This converts human oversight from rubber-stamping into a true review.

The FDD and FDE Symbiosis: Two Halves of the AI Brain

I certainly believe an FDD will work hand-in-hand with an FDE, operating as an inseparable two-person vanguard within the client organization. The sheer complexity of the enterprise environment and the frontier nature of the technology demand it. There’s too much to know about implementing highly secure, performant, latency-optimized solutions with the next generation of AI models for a UX or service design specialist to be the best at bringing the designs to production life.

Consider the technical realities of deploying AI at a Fortune 500 company: navigating massive data privacy compliance frameworks, managing token limits and inference costs, fine-tuning open-source models for highly specific proprietary data, architecting complex vector databases, preventing prompt injection attacks, and ensuring high-availability enterprise infrastructure. These are deep, rigorous engineering challenges requiring world-class talent. An FDD cannot, and should not, be expected to manage them.

There’s no shame for an FDD to need an FDE, just as there should be no shame for an FDE to need an FDD. User research (and the associated systemic design solutions) and technical implementation are fundamentally separate, full-time skills that require different talents and experience. One defines what the system must achieve and how it will impact the human workforce; the other engineers the computational engine to make it a robust reality.

FDD and FDE roles require different skillsets and different mindsets. The two together are unbeatable and can explore anything.

However, AI tools absolutely allow people with each of these two skillsets to blur the lines and dabble in the other’s domain. A UX person can create his or her own functioning prototypes without waiting for a developer. Leveraging AI coding assistants, natural language programming, and advanced no-code environments, an FDD can spin up functional “Wizard of Oz” prototypes to test a prompt chain with users in real-time to validate an assumption instantly.

Engineers should also observe users and prototype solutions themselves. FDEs should actively participate in sitting in on contextual inquiries, bringing their intimate technical intuition directly to the front lines to suggest possibilities the designer might not know exist. This cross-pollination shrinks the gap between idea and execution, fostering deep mutual respect and a shared language.

But when it comes to enterprise-grade transformation, dabbling is not enough. The stakes are too high. The best, most paradigm-shifting research insights, the most radical design solutions, and the most robust, secure, and scalable implementations leveraging the absolute bleeding edge of next-generation AI models will require dedicated specialists working in symbiosis.

When DeployCo or any forward-thinking consultancy embeds a team, that team must be a unified pod. The FDD discovers the right thing to build, maps the organizational psychology, uncovers the true Job-to-be-Done, and architects the blueprint for the new AI-native workflow. The FDE takes this blueprint, selects the optimal combination of foundational models, vector databases, and security protocols, and engineers the system to survive the rigors of reality. Together, they act as the co-founders of the client’s new operational reality.

But even the perfect FDD/FDE pod will fail without an operating sponsor who has the authority to retire the old workflow. Enterprise AI adoption is not primarily a training problem. It is a permission problem. Employees will continue doing the old work in parallel until leadership explicitly removes obsolete steps, changes metrics, rewrites job descriptions, and protects teams from being punished for trusting the new system. Transformation begins when the organization is allowed to stop performing the rituals that AI has made unnecessary.

Even the best FDD/FDE duo will fail unless they secure an executive sponsor at the client who opens up the door to change. This again means that the persuasion skills to convince executives are one of the main job requirements for this job.

Do It Again

The FDD role will likely be long-lasting, because redesigning corporate workflows for AI is not a one-time deal. Whatever new AI workflows are designed now will have to be redone more aggressively in a few years, and then redesigned even more aggressively after that, as AI capabilities keep improving indefinitely as we build more compute.

Some AI influencers have claimed that the fact that OpenAI (and Anthropic, Microsoft, and other AI vendors) are building strong FDE practices means that they don’t really believe in superintelligence. I think that’s an erroneous conclusion. First, we don’t have ASI yet and don’t expect it to arrive until around 2030. Companies need to take advantage of AI now, and anybody who waits until 2030 to redesign their workflows for AI will likely be out of business long before 2030 arrives. Any redesign done now will have to be led by human FDD and FDE staff.

Keep improving. The work is never done, even after launching a redesigned workflow.

Second, even when we get superintelligence, that doesn’t mean that AI will do all the work and that humans will be relegated to the unemployment lines. There will be more work as the economy expands beyond belief, and we keep coming up with new things we want done. Human desires are infinite. There will also be a plethora of new companies founded to serve all these emerging human needs, and all these new companies will need AI workflows designed.

Even when we get artificial super-duper-intelligence (ASDI 😏), we will still need humans in FDD/FDE jobs, because humans will retain two roles that AI is ill-suited for: agency and persuasion/manipulation. Pushing a project forward requires human agency and impatience, and getting other humans to agree requires persuasion and manipulation. Having a physical body that can sit in a meeting room and look executives in the eye while saying “you need to do this,” will remain a human capability. Probably forever: Even though there are certainly initiatives in the works to build robots that are indistinguishable from live meatware humans, other people will still know who’s a human and who’s a robot and will be more likely to be persuaded by a human presence.

For the short-term future (maybe 10 years or so), it’s also likely that the field research I was recommending above should be conducted by human UX anthropologists and not by robots. After that, people may not care whether they are being observed by a living creature or a machine. In fact, I think it’s likely that many workers will prefer field research to be conducted by machines, since it’s less embarrassing to goof up in front of an AI than in front of a human, who will be seen as judgmental, not matter how many times we say, “we’re not testing you, we’re testing the system.”

Conclusion: Redesigning the Future of Work

OpenAI’s massive $4 billion investment in DeployCo and the integration of Forward Deployed Engineers into legacy consulting firms prove that the era of theoretical AI is over. The era of gritty, highly lucrative operational enterprise AI deployment has officially begun.

But as we embark on this transition, consulting giants and enterprise clients alike must avoid the temptation of using revolutionary technology to simply optimize the past. If we just throw brilliant engineers and advanced AI models at outdated, clunky legacy workflows, we will undoubtedly achieve isolated 40% efficiency gains. But those gains will be lost in the friction of broken, human-centric systems, and we will have squandered a once-in-a-generation technological leap.

To truly unlock the exponential economic and creative potential that artificial intelligence promises, organizations must be willing to engage in radical, blank-slate workflow redesign. We must look at our businesses not as a collection of human tasks to be automated, but as a series of business outcomes to be achieved in entirely new ways. We must discard the legacy systems built around human limitations and design new, agentic ecosystems.

This monumental task cannot be accomplished by engineers alone, nor by traditional designers pushing pixels in remote offices. It requires a new breed of practitioner who can bridge the gap between human psychology, complex service systems, and frontier technology. By establishing and championing the discipline of the Forward Deployed Designer, DeployCo and the broader tech industry can ensure that the systems of tomorrow are built on a foundation of deep human understanding and visionary design. The future belongs not just to those who can code the AI, but to those who can design the entirely new ecosystems in which the AI lives.

AI redesign shifts UX value from task completion to enterprise control loops, including authority, accountability, incentives, and operational drift. FDDs will play an important role in driving this transformative AI revolution across business and the global economy. Will you be one of them?

There’s a bright future for those who can redesign workflows for AI and break away from companies’ dependency on legacy human workflows.

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 (30,387 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: Video about scope creep | Demis Hassabis on the present and future of AI | Humans and AI have different interpretations of words to denote probabilities | Very different reasons why people use AI versus the Internet | Google’s disappointing AI announcements

UX Roundup for May 25, 2026 (GPT-Images-2)

Scope Creep Video

I released a new short video about scope creep (YouTube, 2 min): the tendency to keep adding requests to a design project that spoils any clean interaction architecture. Featuritis is a usability disease!

Scope creep, the video. (GPT-Images-2)

Featuritis is a usability disease. Scope creep is common in product design, but adding more features often degrades the user experience. Less is More! The scope-creep scenario in my video has a happy ending, but sadly, most companies end up with bloated software instead.

I made this video with two different animation styles, which are mixed throughout, alternating styles for each new scene:

• 2-D character animation

• Photorealistic

Which style do you prefer? Let me know in the comments.

Ironically, using two animation styles in a single video is an example of scope creep in its own right, but I thought it was fun as an experiment, and I create these videos mainly to test out the steadily advancing capabilities of the AI models.

I reused three of the character designs from my comic strip about my predictions for 2026, which were in a 2D style. I then uploaded the character reference sheets for these characters to the GPT-Images-2 model and asked it to generate new photorealistic character sheets for them, as if they were real people. I am impressed with how well it brought the cartoon characters to life, and in fact, after creating video clips based on the two sets of character reference sheets, I tend to prefer the photorealistic parts of my video.

One of my character reference sheets, in the original comic-strip 2D cartooning style (top), was transformed into photorealism (middle) and transformed again into Pixar-style 3D animation (bottom). I didn’t use the last version because I ran out of AI credits and didn’t want to complicate my final video by switching between three styles. (GPT-Images-2)

I probably still prefer the 2D cartooning style for comic strips.

It is one of the characteristics of AI as a native media form that one can transform the content in various ways, such as changing the visual style. I didn’t do this, but of course it’s also possible to translate the language and have updated lip-syncing, making it look as if the characters speak native, say, Danish or Korean. After one or two additional generations of AI advances, it will also be possible to update the sets and costumes in real time to reflect more culturally appropriate localization for each individual viewer.

AI can instantly remix a video into different languages and different settings. For now, this needs to be done before publishing the video, but in a few years, these adjustments will happen in real time based on each viewer’s individual preferences. (GPT-Images-2)

I used Seedance 2.0 for this project, since it remains the best video model on the market. It’s expensive, though: I ran through a full month’s worth of AI credits (roughly $200) to generate the final two minutes of useful footage. One problem is that even though Seedance generations usually look good (especially when working from attractive reference images), they often don’t sound good. This Chinese model often garbles English words. For example, I had to run through around 10 expensive generations to have it pronounce the word “readiness” correctly. Also, even though Seedance has decent physics, it doesn’t work from a full world model, so in multi-shot clips, when it cuts between camera angles, people often move around and don’t sit in the same place from shot to shot, or the direction in which they are looking suffers continuity errors.

Getting the words pronounced right is currently the biggest problem with Seedance. It also has a tendency to add extraneous words that are not in the specified dialogue. (GPT-Images-2)

Continuity problems permeate multi-shot generations with current video models. For a commercial project, you have to bite the bullet and spend more on a large number of rerolls. For my projects, which make no money, I let many such smaller problems slide and stop when the video is “good enough.” (GPT-Images-2)

Even though the underlying video model for Seedance seems to have remained the same since the version 2.0 release, the sound model must have been tweaked, as its English pronunciation keeps improving. Two months ago, it never pronounced the word “newsletter” correctly. Now, it pronounces it “news-later” about 20% of the time, which is hilarious, but it gets the word right most of the time.

Much as current limitations annoy me, Seedance is constantly improving. As is most AI. (GPT-Images-2)

Demis Hassabis on AI: Real Progress, Real Gaps, No Magic Required

The current AI paradigm of pre-training at scale, RLHF, and chain-of-thought reasoning is not a dead end. According to Demis Hassabis, co-founder of Google DeepMind and 2024 Nobel laureate, these components will form part of any future AGI architecture. The notion that we will discard them in two years and start over is, in his words, implausible.

Demis Hassabis is one of the world’s smartest people. It’s worth listening to his predictions and vision for how AI will help humanity. (GPT-Images-2)

But the architecture is not finished. In a recent conversation with Y Combinator’s Garry Tan (YouTube, 41 min.), Hassabis identified three concrete gaps: continual learning, long-term reasoning, and memory. He places roughly 50/50 odds that one or two more major conceptual breakthroughs are still required before we reach AGI. His personal timeline: around 2030.

The memory problem illustrates the wider issue. Today’s systems shove everything into the context window. Hassabis calls this “duct tape.” A million-token window sounds enormous until you try to record live video, at which point it covers about 20 minutes (only half of the interview on which I based this news item). Brains do not work this way. They retrieve selectively. AI systems do not yet do this well.

The same diagnosis applies to reasoning. Today’s models can solve gold-medal International Mathematical Olympiad problems while flunking elementary arithmetic when phrased oddly. Hassabis calls this “jagged intelligence.” Watching Gemini play chess, he sees the model identify a blunder mid-thought, fail to find a better move, then play the blunder anyway. A precise reasoning system should not do that. The gap is not raw capability. It is introspection: an honest assessment of one’s own thought process.

Agents fare similarly. Hassabis agrees they are hyped. He also agrees they are just beginning. Most current deployments are experiments dressed up as products. Real value is starting to appear only in the last few months. He has not yet seen a vibe-coded game top the charts. He has not seen a teenager ship a hit title using these tools (like he did at age 17). That absence is informative. Craft, taste, and human originality still matter.

This connects to Hassabis’s deepest unresolved question. AlphaGo’s Move 37 was novel. But could a system invent the game of Go itself from a high-level description? Today’s models cannot. The leap from solving problems to generating problems worth solving remains unsolved, as does his “Einstein test,” of whether an AI trained on knowledge through 1901 could derive special relativity.

For builders, two practical observations. First, deep tech is no harder than shallow tech. It is differently hard. Pick problems whose absence would actually matter. Second, if AGI arrives around 2030 and your project takes ten years, AGI will arrive in the middle of your work. Plan for that. The likely architecture is general-purpose models orchestrating specialized tools like AlphaFold, not one giant brain doing everything. Build accordingly.

Smaller models will continue absorbing frontier capability through distillation, with no theoretical ceiling in sight. Inference will not become free; Jevons paradox guarantees demand scales to consume any supply. Open weights matter, particularly on edge devices where local execution serves privacy and latency.

In summary: real progress on AI, still real gaps, no magic required to reach AGI.

Summary of Garry Tan’s interview with Demis Hassabis. (GPT-Images-2)

How Likely is “Likely”? AI and Humans Disagree on Probability Words

Communication relies not just on words but on shared expectations about what those words mean. When a doctor tells a patient a treatment is “unlikely” to cause side effects, both parties assume a certain probability that may not be the same. But what happens when an AI interpreter sits in the middle? In their study, Mayank Kejriwal and colleagues from USC’s Information Sciences Institute posed this question. They compiled a list of 17 probability phrases, ranging from “almost certain” to “extremely unlikely,” and asked both human participants and several leading large language models to assign numerical probabilities to each term. The goal was to see whether machines and people share a common mental scale for uncertainty.

The results reveal a notable disconnect. For extreme phrases like “impossible” and “certain,” both humans and models aligned fairly well, mapping them to very low or high probabilities. But mid‑range likelihood terms like “likely,” “possible” and “maybe” diverged significantly. Humans tended to interpret these words as conveying 50–70 % confidence, whereas the models leaned toward 70–90 %. The discrepancy was largest for “likely”: the average human estimation was about 66 % while GPT‑4 and its peers interpreted it closer to 80 %. Interestingly, more advanced models did not converge toward human interpretations; some even amplified the differences.

Humans and AI interpret common words that indicate a degree of likelihood differently. (GPT-Images-2)

Why does this matter for user experience? Many AI applications, from medical diagnostics to weather forecasting, use natural language to communicate risk. This is probably (ha!) necessary, since most humans are notoriously incompetent at understanding probabilities, so resorting to formal statistics is a doomed idea.

If a chatbot tells a user that something is “unlikely,” the person may assume a lower probability than the model intended or vice versa. Misaligned expectations can lead to overconfidence or unwarranted caution. The USC team suggests that AI systems should avoid relying on vague probability words and instead either translate them into numerical ranges or ask users to specify what they mean. In other words, the system needs to “speak the user’s language” (following usability heuristic 2) rather than projecting its own calibration.

From a design standpoint, the study underscores the importance of clarity and transparency in AI outputs. Interfaces that present probabilities should allow users to drill down into the underlying data or choose their preferred form of expression. Designers should also be aware that training data influences how models interpret language; if a model has been exposed to contexts where “likely” means “almost certainly,” it will carry that bias into other domains. Testing AI systems with diverse user populations can help surface these mismatches and inform correction strategies.

The study invites a broader reflection on how humans and machines align on meaning. As AI becomes ubiquitous, miscommunications may erode trust and hinder adoption. Proactive design choices, like prompting for clarification or offering explicit probability ranges, can mitigate these risks. The research therefore contributes to the ongoing conversation about making AI outputs not just accurate but interpretable and aligned with human mental models.

As always, more research is needed: in particular I would like to see experiments with better ways of illustrating and communicating probabilities to people with different levels of statistical numeracy, recognizing that 90% of the population is completely incompetent at statistics, and only 1% have a decent understanding of probabilities and other stats concepts. After all, cave age people didn’t stop to calculate the probability of being eaten by a saber-toothed tiger. If they saw one, they ran, or they would not have passed their DNA down to us.

Why present-day humans are bad at probabilities. (GPT-Images-2)

Why People Use AI vs. the Internet

GWI (formerly GlobalWebIndex) is a global consumer research platform covering 50 countries. Their survey data shows a substantial difference in the top reasons people use the Internet versus AI.

For both services, the number-one reason is “Finding Information,” mentioned by 60% of Internet users and 59% of AI users. The same number, given the inherent uncertainties in self-reported data.

The following reasons differ dramatically. Main reasons (after finding information) for using the Internet:

1. 58% say “staying in touch with friends and family”

2. 53% say “watching videos, TV shows, or movies”

3. 51% say “keeping up to date with news and events”

4. 49% say “researching how to do things”

5. 45% say “accessing and listening to music”

Basically very passive use cases, except for researching how to do things. There was even a fairly popular survey response where 44% of Internet users gave “filling up spare time” as one of their main reasons for using the Internet. Given how little time we have on this Earth, I find it depressing that many people want to spend it on nothing.

Answers were very different for the main reasons given by AI users (again, after finding information):

1. 38% say “get advice on problems”

2. 37% say “learn or improve skills”

3. 34% say “create or enhance images or video”

4. 33% say “save time or increase efficiency”

5. 30% say “create or enhance written content”

AI use is dramatically more active and creative than Internet use, which is dominated by the passive consumption of content created by others, mainly as entertainment.

The Internet has become a pacifying medium, whereas AI encourages active users, as shown by the top reasons people state for the two forms of computer use. (GPT-Images-2)

• Internet = Veg out

• AI = Create and improve yourself.

The traditional Internet has largely devolved into an engine for passive consumption, while AI has emerged as a catalyst for active creation and self-actualization.

When we designed the modern web during the dot-com bubble, my vision (and that of many others) was a highly interactive empowering environment that would spread knowledge and turn everyone into a creator. Instead, the Internet has largely become the ultimate digital couch.

Looking at the top secondary reasons people use the Web, the underlying theme is frictionless escapism: watching videos (53%), reading news (51%), and listening to music (45%). These are entirely “lean-back” activities where the user is simply an audience member. Much like the legacy media like television we were hoping to replace.

The most tragic statistic is the 44% of users who explicitly view the Internet as a way of “filling up spare time.” Human time is our only truly non-renewable resource. The fact that nearly half of the digital population uses the most advanced communication network in history simply to kill time is existentially depressing.

The goal of UX design has changed over the years: During the early years of the web, we tried to empower users and make them better at navigating the web and expressing themselves in this new media form. Then the goal became to monetize passive users and make them buy and scroll endlessly to see more ads. Now, in the AI era, UX design again aims to put users in charge, but this time with an emphasis on intellectual augmentation and creation. (GPT-Images-2)

Why did this happen? It is a feature of the modern web’s business model. The attention economy is funded by advertising, which requires keeping users on a platform for as long as possible. To do this, tech giants built infinite scrolls, auto-playing videos, and algorithmic feeds optimized to remove all cognitive friction. The Internet was designed to make us “veg out” because a passive, sedated consumer is easier to monetize than an active one.

By contrast, the AI usage data shows a “lean-forward” technology (like we wanted the Web to become). If the Internet is a television, AI is a workshop.

Lean-back vs. lean-forward is an eternal dilemma in user experience. When I originally worked on Web usability in the 1990s, the early Web was definitely a lean-forward experience, but sadly the dominant use has turned into lean-back. We get a second chance with AI. (GPT-Images-2)

The verbs dominating the AI usage data require high cognitive engagement: Get, learn, improve, create, enhance.

The Interface Dictates the Mindset: The core reason for this divergence lies in the user experience. The fundamental interface of the modern Internet is the feed; it requires zero input to receive a dopamine hit. The fundamental interface of AI is the intent. You cannot doomscroll a blank ChatGPT or Seedream text box. To get any value out of AI, you must bring an intention, a problem, or a creative vision to the table. This prompt-driven architecture forces the user to be the director rather than the audience.

We may think we design the UI, but the UI also designs us. It’s a duality, where each partner influences and shapes the other. (GPT-Images-2)

Furthermore, AI radically lowers the barrier to entry for creativity. Tasks that previously required specialized software and years of training on drafting polished essays, writing code, generating art are now accessible to anyone with an idea. AI doesn’t just entertain you; it equips you to build.

An infinitely scrolling feed makes us passive, AI activates us. (GPT-Images-2)

Killing Time vs. Reclaiming Time: The most philosophical takeaway from the GWI data is how the two technologies interact with our human lifespan. While 44% of Internet users are actively trying to burn away their spare hours on “nothing,” 33% of AI users are actively using the technology to save time and increase efficiency.

While the Internet frequently functions as a waiting room where we passively pass the hours, AI functions as a personalized tutor and co-pilot. Users are deploying AI to offload digital drudgery so they can free up their cognitive bandwidth for real life. AI users are not looking to escape reality; they are asking for advice (38%) to help them navigate it.

Time is the ultimate scarce resource that runs out for all of us in the end. That’s why I have always been keen on the second usability quality attribute of efficient use. AI helps us get more out of our limited time. (GPT-Images-2)

The Future: Will the “Create and Improve” Era Last? Right now, the data reflects a beautiful, optimistic moment in digital history. It proves that humans do possess an innate drive to learn, build, solve, and improve themselves: they just needed a tool that scaled to their ambitions without getting bogged down by the attention economy.

However, there will inevitably be immense commercial pressure to integrate AI into passive feed mechanics. If tech companies use generative AI to create infinitely personalized, hyper-addictive streams of entertainment designed to keep us scrolling, AI could become the ultimate “vegging out” machine.

To ensure the data in future reports continues to reflect this current “create and improve yourself” paradigm, we must consciously fight to keep AI a tool of human agency. We must use our machines not to turn our brains off, but to turn our potential on.

The future doesn’t happen to us. We make it happen. It’s our responsibility, especially in UX, to create a positive AI future. (GPT-Images-2)

Google’s AI Announcements Were Disappointing

I was very disappointed with Google’s AI announcements at last week’s I/O conference. Very modest advances, compared to what we’ve been getting used to from the other AI labs this year, where AI has been accelerating and is approaching RSI (recursive self-improvement).

Google I/O was a disappointment. (GPT-Images-2)

To start on a positive note: It is great that Gemini 3.5 Flash is 4x faster: speed is a usability benefit in itself. (I wrote my first article on AI response times in August 2023; that’s how fundamental speed is to usability.)

Speed drives usability by allowing users to stay in a flow state. (GPT-Images-2)

So far, I’ve gotten less impressive results from Gemini 3.5 Flash Extended Thinking than from Gemini 3.1 Pro Deep Think, so I doubt I’ll use it much. (If I want a very simple question answered really fast, I’ll use ChatGPT 5.5 Instant or Grok Fast.) For example, I asked both models for critical feedback on the manuscript for an upcoming article, and Gemini 3.1 Pro Deep Think was more insightful, though GPT 5.5 Pro Extended Thinking was even better.

Google’s new video model is also a very limited upgrade, compared with what we have had from Chinese video models for months. Gemini Omni Flash can generate 10 seconds clips, which is an extremely modest upgrade compared with the 8-second clips from Veo 3.1. In contrast, Seedance 2.0, Kling 3.0, and HappyHorse 1.0 can all make clips that run 15 seconds: those extra 5 secs are an eternity in video production. (Kling even generates in 4K natively, which produces immensely better-looking video than the puny 720p delivered by Gemini Omni Flash. Seedance and HappyHorse aren’t 4K yet, but at least they already generate in 1080p natively and will surely be 4K soon, given the cutthroat competition in China.) Even Wan 2.7 and Grok Imagine run to 15 seconds.

Google is falling behind in video generation. (GPT-Images-2)

I made a small compilation of the exact same prompt on Gemini Omni Flash and Seedance 2.0: speedrunning through Chinese history (Instagram, 1 min.). Which model’s clips do you like best?

At least Google seems to have realized it is falling behind in AI capabilities with these releases, as it reduced the price of the Ultra subscription (needed to access Deep Think) from US$250 to $200 per month. Even better, Google introduced a new, cheaper Ultra subscription plan at $100 per month that offers the same benefits as the $200 plan, but lower rate limits: you get 1/4 of the compute allocation from paying half the price, so if you’re a heavy user of Google AI, do stay on the most expensive plan. But I promptly downgraded to $100 since I’m currently using OpenAI and Seedance much more than I’m using Google’s models.

Lower prices for Google AI subscriptions reflect its less competitive performance. (GPT-Images-2)

If Google becomes more competitive later, I can always upgrade back. Even though Google’s subscription UX has been terrible in the past, I should give them props for now having put a little bit of usability effort into separating customers from their money: it’s now easy to find the subscription plans and change your level.

Google’s subscription UI has changed from terrible to good. Nice! (GPT-Images-2)

So if you’re a Google Ultra subscriber but don’t use it enough to justify the high-cost plan, downgrade to $100/month. Google does not advertise the availability of this plan widely: they sent me an email touting the price reduction from $250 to $200, but without any mention of the $100 option. Borderline unethical customer communication.

Google has also reverted to its old bad habits, with a sprawling, confusing UX architecture for its AI offerings that are scattered ever more widely. For one, Google’s video generation changed its name from Veo to Gemini Omni. Unnecessary brand confusion right there. Second, Google is launching a new image editing capability any day now, called Google Pics, that will seemingly not be integrated with Google’s main image-generation feature at Google Flow. Confused already?

The architecture and naming schemes conspire to create a highly confusing user experience for Google’s universe of AI offerings. (GPT-Images-2)

From the demo, Google Pics looks promising, offering object-level editing similar to that Reve has had for ages. But why does this image editing feature have to be moved into a new location rather than being integrated with image generation?

My prediction number 6 for 2026 was that Google would get its AI act together. Sadly, it seems I was wrong. (Though they still have half a year to turn the supertanker.)

Will the Google behemoth be able to mend its ways and have a coherent AI strategy with a clean user experience? They still have time, but need to take this problem more seriously. (GPT-Images-2)

UX Design Video Game Launch Poster

What if UX design were a video game? Here’s the launch poster for this game. Somebody should build this. (Prompt credit: LudovicCreator.)

(GPT-Images-2)

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 (30,387 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 not merely making designers faster at producing screens, prototypes, and research summaries. It is changing the object of design itself. The UX profession’s most valuable contribution stops being UI production and becomes the design of intent itself: defining what good means, encoding judgment into live systems, setting boundaries for agentic behavior, maintaining coherence across many parallel outputs, and preserving human purpose when software starts to act rather than merely wait for commands.

(All my images in this article were made with GPT-Images-2)

Except for die-hard holdouts in denial of the real world, most UX professionals understand that the profession has fundamentally changed as AI takes over lower-level tasks. Humans are left to orchestrate the AI, but will shortly stop performing traditional craft-related activities, such as UI design and user research. AI moves the design problem to a higher level of abstraction. The work is shifting from drawing artifacts to defining intent, taste, constraints, behavior, coherence, and responsibility across human-and-agent systems.

AI uplevels the responsibility of human UX professionals: they will not design the user interface, but the higher-level parts of the user experience.

The fundamental nature of the change in the UX profession crystallized for me during a recent workshop titled “Design Futures Assembly,” which Jeff Veen organized for a small group of UX leaders. Participants included the UX chiefs for most major AI labs, heads of UX for several major software vendors and Internet properties, a few UX thinkers like myself, and a couple of deans from top design schools. An impressive group.

The ground rules for the workshop were intended to encourage frank discussions and prohibited quoting anybody by name. Thus, I’m only naming the workshop organizer, Jeff Veen, who has had an illustrious career, including the founding director of interface design for WIRED magazine and its pioneering website HotWired (which interviewed me in 2000), UX Director at Google, Founder of Typekit, VP at Adobe, and even one of my invited speakers back when I ran a usability conference.

From Artifact to Intent

We know that computer–human interaction is becoming intent-driven, as AI shifts the user’s role from operator of a UI to supervisor of the AI doing the work. The exact same change is happening for UX design. Not really a surprise: why should design be different than any other job?

Users and designers both experience the same fundamental change in knowledge work: humans are no longer operating the computer to directly produce work; instead, we supervise the AI that’s doing the work.

UX is not exempt from the same abstraction shift that is reshaping every other knowledge profession. The difference between users and designers is that the intent itself becomes a designed object.

But there is an important distinction between intent as a user input and intent as a design goal. A prompt is not intent. A prompt is merely the user’s first rough attempt to externalize intent. Real intent includes the desired outcome, the unstated tradeoffs, the user’s tolerance for risk, the context in which the result will be used, the standards by which success will be judged, and the failure modes the user would find unacceptable.

A prompt is not the user’s true intent. We must dig deeper.

AI turns design from an artifact-production discipline into an intent-shaping discipline. The designer’s new job is not personally making every screen, but rather defining what good means, encoding that definition into systems, steering agents and collaborators toward it, and knowing when speed produces value rather than merely volume.

Design jobs will not be saved by defending old tasks. Decels who stick to the old ways will be unemployable in 5 years. Design jobs will be saved by taking responsibility for the new design material. AI will generate, prototype, code, adapt, personalize, conduct user research, and analyze the findings. The UXer’s job is to make sure all that new power has direction, coherence, and human purpose.

Chat is only one possible input/output pattern for AI use. It is not a complete product paradigm, and likely not the most important way we will be interfacing with AI in the future. The richer opportunity is to design systems where intelligence changes the shape of the interface, the sequence of work, the division of labor, and the user’s relationship to the task.

Chat was the first popular UI for AI, but is highly limiting, so we’ll need a richer world of AI UX.

Our new design problem is behavior design for systems. Designers are no longer specifying every path through a product. They are defining the rules, boundaries, tools, permissions, escalation points, and feedback loops that let humans and agents work together productively. That is a much larger design canvas than screen layout.

“Taste” vs. “Opinion” as Human Moats

A crucial part of providing human direction for AI involves drawing a sharp distinction between “taste” and “opinion.” Many designers mistakenly believe their refined taste will serve as a permanent moat against automation. This is a dangerous illusion. AI models are exceptional pattern matchers and will soon commoditize UI taste just as they have commoditized boilerplate code.

The true, durable high ground is opinion: a distinct, assertive point of view about what actually matters in the real world and what is genuinely worthy of a user’s increasingly scarce attention. When generating infinite product variants becomes practically free, the core skill shifts to editorial coherence. Success means knowing which of the million possible AI-generated iterations to ship, and which to aggressively undo, ensuring the product still feels like it was crafted by a single, deliberate mind rather than a committee of autonomous agents.

Opinion is our judgment according to our values. Taste will be AI’s domain. Opinion will be human. (My apologies to Danish artists P. S. Krøyer, Anna Ancher, and Michael Ancher for making their paintings the subject of this art critic’s opinions.)

The word “opinion” implies that there is no one truth, but that has always been the case for UX. Many solutions can all help users, and it has always been UX’s job to identify a solution that balances multiple stakeholders’ requirements with customer needs. The very fact that something is an opinion means that it’s well-suited for humans to defend, based on our evolutionary animal nature and inherent drive to impose our will on the world.

When production is scarce, organizations reward the person who can make the thing. When production becomes abundant, organizations need the person who can decide what is worth making. This is the editorial turn in design. The designer becomes the person who can say: this variant is fluent but wrong; this workflow is impressive but unnecessary; this automation is fast but disrespectful of the user’s need for control; this generated interface is attractive but has lost the company’s voice; this agent completed the task but taught the user nothing and left no recoverable trail.

The danger is not that AI will produce a few bad screens. It will certainly do that at first, but less so each year. In any case, bad screens are an old problem that we know how to defeat through user testing. The new danger is that AI will produce many adequate screens that all seem defensible in isolation and incoherent in aggregate. Mediocrity will arrive well-dressed. The designer’s role is to prevent the organization from drowning in plausible options.

This is why opinion becomes more valuable, not less. Taste chooses the best alternative design, and AI will do that just fine. Opinion chooses what deserves the user’s time. AI will become very good at taste, because much of taste is pattern matching over precedent. Opinion is harder. It requires a point of view about what matters, what should be avoided, what should be simplified, what should be refused, and what kind of product the organization is willing to become.

The Source of Truth Is Live Systems, Not Static Design Artifacts

AI also collapses the old separation between design mockup, prototype, and production. In the new world, designers increasingly work in or near the codebase, where prototypes are branches, where agents operate against real components, and where the design system becomes machine-readable infrastructure rather than a static library.

This changes the politics of design work. A Figma frame used to be interesting because it represented a suggested future. A working pull request, live branch, or cloud-deployed prototype is more persuasive because it is a future version of the product. Design communication becomes less about handoff and more about direct participation in the product substrate.

AI allows us to prototype a rapidly branching set of design ideas, all of which become live and can be tested with users.

The old unit of design work was the artifact: the sketch, the wireframe, the mockup, the prototype, the research report. The new unit of work is increasingly the workspace: a live environment that contains the code, components, prompt instructions, product requirements, research context, analytics, brand rules, agent roles, and critique history needed to move a product forward.

This is a major conceptual shift. A file is something you hand off. A workspace is something you inhabit. A file asks others to interpret your intention. A workspace carries intention forward, because the relevant context travels with the work. It lets a designer, engineer, content designer, product manager, and agent swarm operate on the same problem without reducing collaboration to a sequence of screenshots and Slack messages.

Store context in a single workspace that is shared among many AI tools and all human collaborators.

The design challenge is therefore no longer just the interface inside the product. It is also the interface around the work. How does a team see what the agents changed? How does it compare ten parallel explorations without choosing the flashiest one? How does it preserve rationale? How does it know which constraints were active when an output was generated? How does it recover when a promising branch goes wrong? How does it avoid two agents solving the same local problem in ways that create a global inconsistency?

This is where design operations and product design merge. The workflow surface becomes part of the product’s quality system. A poor workspace produces poor design, even when the individual outputs look good, because decisions lose their lineage. Future design maturity will depend on whether organizations can make intent, context, and critique persistent enough for both people and agents to use.

The traditional cross-functional handoff is dead. But while killing the handoff removes a major bottleneck, it also destroys the shared understanding that the friction of old processes naturally forced upon us. Today, a designer with an AI can bypass development, and a developer with an AI can bypass design. To prevent organizations from devolving into parallel chaos, the new collaborative primitive must become the shared, machine-readable workspace. These workspaces bundle an isolated copy of the live codebase with the team’s codified positions, such as brand voice, testing conventions, and accessibility rules, all version-controlled together. Only by explicitly encoding a team’s collective opinions into the very prompt layer of these operational workspaces can swarms of human and AI agents operate simultaneously without stepping on each other’s toes.

Making Design Intent into Machine Objects

The most important implication is not that “designers should code.” The deeper implication is that design intent must become operationally legible to machines. Design systems, brand rules, research insights, accessibility standards, content voice, and interaction principles need to be encoded in forms that agents can use. Otherwise, AI will faithfully accelerate incoherence.

The design system thus stops being a component library and becomes an operating system for taste. Tokens, components, and usage rules are only the visible layer. Underneath must be a deeper set of instructions about brand behavior, interaction philosophy, accessibility standards, motion logic, content tone, escalation patterns, and product judgment. The system must know not only which button to use, but when not to add a button at all.

This is a major upgrade in the status of design systems. In the pre-AI era, many design systems were partly theater: useful for consistency, but often bypassed under deadline pressure and treated as documentation rather than infrastructure. In the AI era, a weak design system becomes dangerous. It gives agents a bag of parts without a theory of assembly. The result is not chaos in the old sense, but a smoother, faster, more scalable form of inconsistency.

Machine-readable does not mean machine-owned. Humans still decide the principles. But principles that remain only in slide decks, critique conversations, or the memories of senior designers will not survive contact with agentic production. They must be encoded, versioned, tested, and updated. In other words, design taste needs an API.

This also suggests a new form of design debt. We used to accumulate design debt when teams shipped inconsistent components or patched over poor flows. Now we will accumulate intent debt: undocumented assumptions, vague brand guidance, missing escalation rules, untested agent permissions, and research insights that never become usable by the systems doing the work. Intent debt will be harder to see than visual inconsistency, but it will be more damaging because it compounds invisibly through every generated output.

Intent debt will accumulate unless we’re careful.

No Single AI Tool Stack

The sheer velocity of AI adoption has completely outpaced the structural maturity of organizations and educational institutions. The era of standardizing around a single, unified design software stack has ended, replaced by a chaotic “wild west” of bespoke internal micro-tools, parallel agent swarms, and constantly shifting workflows.

While individual practitioners are expanding their technical capabilities at an unprecedented rate, UX teams don’t have a shared understanding of their toolkit. This disconnect leaves teams to navigate the transition in highly unstructured and unpredictable ways.

AI is by no means a settled field: new tools emerge monthly, if not faster, as do upgrades to old AI tools, which change what they can do, and therefore how they should be used. It’s simply too early to mandate a single AI stack and standardize on a small set of enterprise licenses that everybody should use.

The absence of a canonical stack should not be viewed only as a procurement headache. It is evidence that the UX profession has entered an exploratory phase comparable to the early web, when the job had not yet hardened into stable titles, methods, or toolchains. In such periods, premature standardization ties you down. It protects the organization from confusion by protecting it from discovery.

However, “wild west” should not mean “anything goes.” The better model is marked trails through unstable territory. Organizations should allow many tools, but require shared reporting about what was attempted, what failed, what was learned, what risks appeared, and what should be reused. The standard should move up one level: not standard tools, but standard reflection.

This is a subtle but important leadership move. Leaders should not ask, “Which AI tool should every designer use?” They should ask, “How does our team metabolize new capabilities faster than our competitors, without turning every designer into a lonely tool scout?” The answer is a learning system: demos, critique sessions, shared playbooks, internal examples, reusable prompts, agent configurations, and a culture that treats experimentation as communal infrastructure rather than private advantage.

The winning UX organizations will not be those that guess the right tool six months earlier. They will be those that learn how to learn tools continuously.

Currently, many UX department heads at the workshop use similar solutions to AI tool overload: frequent team get-togethers, where people would demo new AI workflows they were experimenting with.

Such lunch-and-learn experience sharing is the right lightweight response for the current stage, but it should evolve into something more systematic. The goal is not to standardize prematurely on one tool stack. The goal is to standardize on what counts as learning. A mature UX organization should have a regular mechanism for asking: which workflows actually improved product quality, which merely increased output, which created hidden cleanup work, and which changed the designer’s judgment for the better?

Which new AI workflows improved the quality of the final output, and which didn’t? Finding out creates strong alpha at this moment where nobody knows.

Evaluation, therefore, becomes part of design. Teams will need AI-era critique rubrics, not just AI-era tools. These rubrics should ask whether an output preserves user agency, whether it reflects the product’s point of view, whether it reduces or increases cognitive load, whether it fits the user’s real context, whether it provides recoverability, whether it has an intelligible rationale, and whether the system would behave acceptably at scale.

We need new evaluation rubrics for AI designs to guide our human opinion on where to tell it to proceed and when to stop it.

Eventually, design critique will be partly automated, but it should not become fully automated. Automated critique can catch violations, compare against standards, and flag likely slop. Human critique must still ask whether the thing is worth doing. That distinction matters. Machines can increasingly enforce the rules, but people must keep revising the reasons.

Role Boundaries Dissolve Faster Than Organizations Can Absorb

Workshop discussions demonstrated a mismatch between practice and structure. Designers are using more tools, shipping more code, blurring into product and engineering work, and collaborating with agents as well as people. But performance metrics, compensation systems, career ladders, hiring criteria, and collaboration rituals have barely caught up.

This creates a strange organizational condition: individuals are moving quickly, while institutions are still calibrated for the previous workflow. Teams can now produce more variants, more prototypes, more code, and more experiments, but they often lack a shared operating model for deciding what counts as good work.

This calibration gap leads to a structural danger: measuring the impact of AI adoption through obsolete proxies. When designers and product managers are evaluated on the sheer volume of features shipped or pull requests merged, AI will gleefully help them hit those targets, masquerading metric inflation as genuine business value. Rewarding output volume in an era of infinite algorithmic generation actively incentivizes the production of digital slop. Organizations urgently need new performance frameworks that discard legacy productivity metrics. Career ladders must pivot to evaluate a practitioner’s ability to orchestrate multi-agent workflows, their editorial judgment in curating high-quality outputs, and their success in designing trust and seamless delegation between humans and intelligent systems. When AI provides infinite scale, the most valuable design skill becomes knowing when to say no.

Loneliness among UX worker bees is one of the most important side effects. AI makes individual designers more capable, but it also pulls work into private loops between one person and a machine. The old cross-functional friction was inefficient, but it created shared understanding. The new design organization must deliberately manufacture collaboration, critique, and shared taste, because they no longer happen automatically through handoff bottlenecks.

UX workers are experiencing loneliness as they are pivoting to judging work rather than creating it, which was usually more collaborative.

Once recognized, UX staff loneliness is solvable. UX leadership needs to create more communication, critique, and team-building, especially through in-person work sessions. The one thing humans can do that AI cannot is to be physically present: embodied animals in a shared room, reading each other’s hesitation, energy, humor, confusion, and commitment. Bringing these human animals together is not nostalgia. It is infrastructure for judgment.

The counterweight is not nostalgia for old tools. It is stronger grounding in process, culture, research, systems thinking, and human judgment. Designers will need more cultural fluency, not less; more ability to critique, not less; more understanding of users and organizations, not less. AI can generate screen designs, but it does not yet know what a suggested step means in a given culture, company, market, or human situation.

AI’s contextual knowledge will obviously grow, so its analysis of the interplay between UI design and organizations and culture will surpass that of humans in a few years. But humans still retain that raw animal superiority, making their contributions to decision-making meetings more valued than even the most accurate data from AI.

Coherence Is the New Scarcity

AI makes design production abundant, but UI abundance makes UX coherence scarce. When a product team could only produce a few design alternatives, coherence was partly protected by scarcity. The same small group of people saw most decisions. The same designers touched most surfaces. The same bottlenecks that slowed the work also forced some consistency.

AI removes those bottlenecks. That is good, but it also removes the accidental coordination those bottlenecks provided. Ten people and twenty agents can now create more work in a day than a traditional team might have created in a month. The risk is not only inconsistency in the visual layer. The deeper risk is a fractured product mind: each feature locally optimized, each flow plausible, each agent doing its assigned job, and the whole product feeling as if nobody is home.

This makes coherence a management problem, a design problem, and a cultural problem at the same time. Coherence is not sameness. A product can have variety, surprise, personality, and local adaptation while still feeling as if it has one mind behind it. Coherence comes from shared intent: a common view of the user, the product promise, the acceptable tradeoffs, the brand’s behavioral character, and the product’s theory of value.

In the AI era, the designer becomes a guardian of this product mind. That does not mean personally approving every output. That would recreate the old bottleneck in a more exhausting form. It means creating the conditions under which many contributors can make local decisions that still add up to a coherent whole.

Research Is Fuel for Agentic Systems

User research also changes when design moves from artifacts to intent. The traditional research deliverable was often a report, a deck, a journey map, or a set of recommendations. These artifacts were already lossy. They depended on busy product teams reading them, remembering them, believing them, and applying them later under deadline pressure.

In an AI-mediated design process, research must become operational context. Findings should not merely be summarized for humans. They should be structured so agents can use them: user goals, vocabulary, misconceptions, decision criteria, emotional triggers, accessibility needs, domain constraints, observed workarounds, and unresolved questions. A research repository that cannot be queried by agents will become as obsolete as a design system that cannot be used by code.

Once user research findings become part of the AI’s background repository, they can drive agent behavior.

This does not make researchers less important. It makes their judgment more important. AI can summarize transcripts, cluster observations, and generate plausible themes, and soon do so better than 90% of human user research staff. But the human’s distinctive contribution is converting research insights into business profits by driving change across the organization, despite inevitable resistance.

The goal is not faster research reports. The goal is a living model of users that improves the behavior of the product. When agents design, test, personalize, or act on the user’s behalf, they should be drawing on the best available understanding of users, not on generic assumptions scraped from the Internet.

The Future of UX

The next generation of UX professionals will not be saved by proving that humans can make better mockups than machines. That is a lost cause, as AI gets better every year while humans don’t. They will be saved by taking responsibility for the terms under which intelligence enters the product. That includes intent, permission, evidence, escalation, recoverability, coherence, culture, and taste. It includes knowing when speed is a gift and when speed is a way to avoid thinking.

The UX profession is not becoming smaller because it’s no longer protected by the old monopoly on designing good-looking artifacts. The design surface has expanded from the screen to the system, from the flow to the agent, from the component to the rule, from the prototype to the live branch, and from the handoff to the workspace. The work is harder to see because it is no longer conveniently contained inside a rectangle. But it is still design: the shaping of what people can do, understand, trust, and become through technology.

It’s OK not to sweat the small stuff anymore (let AI handle it!). Let’s embrace a broader view of the UX profession.

Our new goal: shape intent, don’t design low-level UI artifacts.

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 (30,387 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: Case study of redesigning workflows for AI in computer game studios | Can you live 4 days without AI? | AI increases learning when used as a tutor and hurts when doing the exercises for the students | A streaming media platform for your job | DeepMind demos integration between voice input and pointing

UX Roundup for May 18, 2026 (GPT-Images-2)

Gaming Proves It: Task-Level AI Is a Dead End

A new Wharton research paper confirms what I’ve been arguing: bolting AI onto existing workflows produces negligible organizational impact. The real gains come from redesigning how work gets done.

Zimran Ahmed studied 20 computer game studios from AAA behemoths to scrappy indie teams and found a pattern that maps almost perfectly onto the framework I laid out in my analysis of redesigning workflows for AI. Studios that treated AI as a task-level productivity tool saw individual benefits that never aggregated into business results. Studios designed around AI from the start saw cycle times collapse by 4–20x.

The gaming industry is the ideal test case because it has used AI longer than most sectors. If any industry should have an easy time adopting generative AI, it’s this one. But organizational barriers were identical to those in every other industry: tacit knowledge, siloed teams, and human resistance to change.

Copy-and-Paste AI Is Local Search by Another Name

Ahmed describes the universal first step: give employees access to ChatGPT Enterprise and let them experiment. He calls this “Copy-and-Paste AI” where the user puts context into a chat window, gets a result, and pastes it back into their work. Every studio started here. It required no organizational change. And it produced no organizational impact.

Task-level use of copy–paste AI wins a few points, while the big prize of complete workflow redesign for AI flies away. (NotebookLM)

This is precisely the local search bias I discussed based on the earlier INSEAD/Harvard data. When you hand people a tool and say, “Be more productive,” they optimize the task immediately in front of them. A product manager drafts better emails. A designer generates placeholder art. Useful, but irrelevant to firm-level throughput. In the INSEAD/Harvard experiment, the control group did exactly this, and generated only half the revenue of firms that redesigned whole workflows for AI.

When you focus on your own local problem, you don’t know what you’re missing in the larger context. (NotebookLM)

The Mapping Problem Lives in Tacit Knowledge

Ahmed’s central insight about why workflow automation stalled is devastating in its simplicity: multi-person workflows run on unwritten rules, and nobody can automate what nobody has documented. The true process differs from the SOPs. Before AI could replace handoffs, someone had to extract, reconcile, and codify institutional memory — and employees had every incentive not to cooperate, because tacit knowledge is job security.

Unwritten rules (tacit knowledge) must be made explicit for AI workflows to work. This cannot happen if each group hoards its private knowledge. (NotebookLM)

This is the mapping problem. The INSEAD researchers demonstrated that the binding constraint on AI profitability is not technical access or coding skill. It is the cognitive and organizational challenge of discovering where AI creates value across the production system. Ahmed’s interviews supply the human texture behind that abstract finding: artists holding crisis meetings, art directors pledging to “support this, but then protest in other places.”

Boundary-Crossing Is Workflow Redesign in Miniature

The most actionable finding is what Ahmed calls “reaching across boundaries.” Individuals used AI to complete tasks that previously required other teams. A product manager wrote database queries. A finance controller automated data extraction. An engineer generated art assets. These people already possessed the domain knowledge; they just lacked the technical execution skills that AI could supply.

Crossing boundaries led to greater workflow improvements than when employees optimized a local problem. (NotebookLM)

This is bottom-up workflow redesign. Each boundary-crossing instance eliminated a handoff, which is exactly what I recommended: remove handoffs, parallelize variants, move humans to exceptions. The product manager who reduced query failure rates from 80% to 5% by iteratively documenting his tacit knowledge into a markdown file was solving the mapping problem in real time, building what Ahmed calls “Read/Write AI”: a compounding knowledge system.

AI-Native Studios Are Company Redesign Made Real

The paper’s most striking data comes from AI-Native studios built on AI from the founding. Teams of five generalists replaced departments of specialists. A vertical slice that would have taken four months took four weeks. Thirty icons were produced in one day instead of weeks. Ahmed calls this “pipeline collapse.” I call it the inevitable consequence of designing the firm around AI rather than inserting AI into the firm.

A small team of AI-empowered generalists replaced a large number of AI-less specialists in one of these case studies. (NotebookLM)

These studios embody every principle from the INSEAD experiment: they eliminated human glue, parallelized exploration, moved humans to judgment and evaluation, and built living documentation that AI agents could reference and update. They achieved what the treated startups in the field experiment achieved: dramatically more output with dramatically less capital and the same headcount.

The Lesson Is Universal

Even though this was a study of computer game studios, the lessons generalize. Every industry has tacit knowledge trapped in specialist silos, handoffs that consume weeks, and employees who optimize locally because they’ve never seen the whole system mapped. Ahmed’s conclusion matches mine: don’t ask “what can we automate?” Ask “how would we build this if we designed around AI from day one?”

The studios that asked the second question didn’t just make games faster. They made games that weren’t possible before. That is the difference between task optimization and workflow redesign, and it is worth a fortune for the business.

We can all learn important lessons about redesigning workflows for AI from this case study of games studios doing so. (NotebookLM)

(Hat tip to Ethan Mollick for alerting me to this paper from his department.)

As I have mentioned many times, the credibility of research findings increases when the same (or roughly the same) results emerge from different research projects conducted by different research groups, using different methodologies and studying different domains. Here, we have a French research group and an American one agreeing, even though one used a randomized controlled quantitative study and the other used a qualitative interview-based method, and one studied incubator startups and the other targeted computer game studios. I believed the first study because it was big and thorough, but I believe it even more now.

Deprivation Study: Living for 4 Days Without AI

Could you go without AI? Could you do it for 4 whole days?

In user experience, we typically measure what happens when people use a system. We run usability tests and calculate task success rates. However, some of the most profound behavioral insights come from the exact opposite approach: taking the system away entirely.

In user research, this is called a deprivation study. You rarely understand a technology’s true impact until it breaks. When a system functions perfectly, it becomes invisible. To truly see how deeply a tool has embedded itself into human behavior, we force users to stop using it and observe the fallout.

Participants in a Korean deprivation study were asked to do just that. Eunseo Oh and colleagues from the Korea Advanced Institute of Science and Technology (KAIST) conducted a small qualitative study where 10 users volunteered to refrain from AI use for 4 days. Whenever they felt the urge to use AI, they logged their context, emotional state, and workarounds into a diary, followed later by in-depth interviews.

Deprivation studies are a standard user research method to investigate the role of a technology in users’ lives. (GPT-Images-2)

AI vs. Smartphone Deprivation

Deprivation studies are a highly revealing qualitative method. A decade ago, classic deprivation studies of smartphone use yielded striking results. When researchers confiscated users’ smartphones, the immediate result was acute anxiety, “phantom vibrations,” and intense FOMO (Fear Of Missing Out). Eventually, this withdrawal led users to reclaim their present-moment awareness and reflect on their mindless scrolling.

AI deprivation tells a fundamentally different story. While taking away a smartphone forces users to confront their emotional dependencies, taking away AI forces knowledge workers to confront their cognitive dependencies. Losing a smartphone causes emotional withdrawal; losing AI causes severe workflow paralysis and a crisis of professional competence.

The Pain of Withdrawal: New Usability Baselines

The study’s main finding is that AI is no longer perceived as optional software; it has become implicit work infrastructure, much like electricity. When this infrastructure was removed, the researchers observed profound disruptions:

• Traditional Search is Now an “Excessive” Burden: Without ChatGPT, users had to revert to traditional search engines like Google. Participants reported that synthesizing, reformulating, and iterating search keywords felt unexpectedly cumbersome. AI has permanently raised expectations: having to translate human intent into machine-readable keywords and sift through links is now viewed as an unacceptable cognitive load.

Traditional search felt cumbersome without the ability to get a simple, synthesized answer from AI. (GPT-Images-2)

• The Social Cost of Asking Humans: Bereft of AI, users had to ask human colleagues for help. Curiously, they hated it. Accustomed to the infinite patience of a chatbot, participants perceived asking humans for assistance as a heavy social burden. One cited a fear of being viewed as a “finger prince/princess,” Korean slang for someone too lazy to look up basic information.

Without AI, people felt they were imposing on other humans when asking them questions. (GPT-Image-2)

• Lowered Standards and Task Avoidance: Generative AI provides massive productivity gains. But without AI assistance to polish writing or catch bugs, participants actually lowered their standards. Rather than expending manual effort to achieve excellence, they accepted mediocre outcomes to save time. Furthermore, they simply abandoned tasks that felt too daunting without an AI copilot.

Users lowered their standards without AI to help them refine their work. (GPT-Images-2)

The Silver Lining: Reclaiming Professional Value

Despite the friction, the deprivation period yielded a profound silver lining. By forcing users to do the heavy lifting themselves, the study revealed the hidden costs of AI dependency.

When participants were forced to construct and trace their own reasoning from scratch, they reported a much clearer, sustained understanding of their work. More importantly, they reported a renewed sense of ownership. Work produced with AI often feels alienated and “not my own.” Completing tasks manually fostered a sense of pride and professional integrity that AI-delegated work had stripped away.

People felt a renewed sense of ownership in the slower work they produced without AI. (GPT-Images-2)

Finally, because manual work takes much longer, participants were forced to rigorously prioritize. Without AI, users had to carefully evaluate which tasks were actually core to their professional identity, filtering out the busywork.

The conclusion of this research is stark: AI use is now socially inescapable. For UX professionals, the central challenge is no longer whether to use AI, but designing systems that allow users to leverage AI without sacrificing their professional agency.

AI Increases Learning When Used as a Tutor, But Hurts When Doing Exercises for the Students

By now, this is an old insight, but it’s always nice when new research studies from new scientists confirm what we already know: this vastly increases our confidence in the results.

AI in education can help or hurt the students’ learning, depending on how it’s used. Without understanding this basic fact, you will get confused by the many studies that are sometimes favorable for educational AI and sometimes have negative findings.

The basic finding is that if the students use AI to do their work for them, they don’t learn. Students may feel good about finishing assignments and exercises quickly, but the goal of student “work” is not to produce the deliverable. The old saying “the journey is the reward” is the point of student assignments: the student learns by working through the problem, not by getting the solution handed down by AI (or by a human helper).

On the other hand, learning is improved when students use AI as a tutor, to guide them through the material, to explain tough concepts tirelessly in many ways, and to provide hints when they are stuck on a problem.

A new paper by Hamsa Bastani and colleagues from the Wharton School shows this once again. In this case, the researchers studied almost 1,000 Turkish students taking mathematics courses in the ninth, tents, and eleventh grades. Three study conditions: (1) A control group with no AI. (2) Full AI, which they inevitably used to solve the math problems for them. (3) Tutoring AI, which provided step by step hints instead of direct answers to problems.

When doing practice problems, students in both the AI conditions did much better than the control group. But in an exam without access to AI, the students in the full AI group did substantially worse than the control group, indicating that they had learned less, not more, from the more efficient completion of the exercises when AI did most of the work for them. There was no statistically significant difference in this exam between the students using the tutoring AI and the control group, indicating that the problem for the “Full AI” students was not AI as such, but rather the way they used it.

Other studies have also found increased learning on final exams for students with tutoring AI, but even if that was not measured in this study, simply the ability to proceed faster through the curriculum is a benefit in itself.

AI can help or hinder learning, depending on how it’s used. (GPT-Images-2)

Streaming Media Platform for Your Job

What if there were a streaming media platform for your job? How would your home screen look? Here are some examples, drawn with GPT-Images-2.

Are there any of these videos you want me to make? Let me know in the comments.

Put That There, For Real

DeepMind has released a short demo video showing the integration of voice input with mouse-based pointing on the computer screen. This is a very natural interaction technique for certain operations on your computer. For example, “make THIS bigger.”

I am immediately reminded of Richard Bolt’s classic Put-That-There research system from the MIT Media Lab in 1980. The system featured a large projection display showing simple colored geometric symbols and labels that represented objects in a scene. Bolt’s innovation was the explicit treatment of speech and gesture as a single, coordinated communicative act rather than as separate input modes. The name of the system was how you used it: You would speak an action (e.g., PUT), point to an object, and point to where it should be moved.

46 years later, DeepMind’s new demo does roughly the same, but with a much larger set of operations, and allows users to use their own computers and use their own data rather than having to sit in a specially equipped projection room.

Compared to the classic WIMP-driven graphical user interface, where commands and objects required separate interactions to drive an action, the parallel specification of what should be done (through the voice channel) and where that should be done (through mouse pointing) streamlines the interaction by unifying the two in time. This is roughly similar to the way the PenPoint system for the GO pen-driven tablet from 1992 allowed users to unify objects and commands into a single gesture. For example, you could make a word bold by writing a “B” on top of that word. (The “B” gesture specified the command, and where it was written specified the operand.)

Exciting to see continued progress on these pioneering interaction techniques, driven by today’s AI capability to broaden the scope, and thus the utility, of unifying operator and operand specification.

History repeats itself, but better this time around. (GPT-Images-2)

Running Through Time

Just for fun: a short AI video of running through time (Instagram, 40 secs.), where I asked Seedance to generate based on a minimum of direction from me. I think it did quite well.

In general, I prefer to direct my videos more intensely so that they reflect my detailed vision. However, as AI gets better, it’s likely that creative vision will have to uplevel and stop obsessing over the details, which AI models will fill in better than humans could do.

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 (30,387 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: Humans have been dreaming of thinking machines for more than 2,700 years, but real work on building them started in 1943 with the first neural networks. Progress during the following 80 years was slow, with many disappointments and “AI Winters,” but advances built up over time as compute scaled. Themes seen during this long history persist to this day.

I made all the illustrations in this article with GPT-Images-2.

This is the story of artificial intelligence, surprisingly rich and full of twists and turns. In fact, we start with the pre-history of AI, before there were computers, but people still imagined artificial minds.

The 35-page comic book (plus this cover) used to illustrate this article is my most ambitious comics project yet. Much longer than my 20-page retelling of Hamlet or 18-page History of the Graphical User Interface. My narrator characters, Alice and Zimo, span the alphabet and centuries of history, which required many panels.

Fictional Thinking Machines: The Prehistory of AI

The dream of creating artificial intelligence is not a product of the computer age. It is not even a product of the industrial age. It is at least 2,700 years old.

For most technologies, engineers invent a new technical capability, and then UX designers figure out how humans should interact with it. With AI, humanity collectively wrote the Product Requirements Document (PRD) centuries before the first line of code was ever compiled. We spent millennia designing the frontend interface in our imaginations, completely unconstrained by the realities of backend engineering. This period of “Proto-AI” established the impossibly high expectations and deep psychological biases that still plague our AI interaction designs today.

The Robots of Fiction

Since the dawn of civilization, humans have craved an interface that could entirely eliminate the cognitive load of decision-making and physical labor. Greek mythology gave us Talos (first mentioned around 700 B.C. by Hesiod), a giant bronze automaton programmed to protect the island of Crete: essentially an autonomous security drone wrapped in an anthropomorphic shell. The Iliad describes Hephaestus’s golden mechanical servants who could speak and reason. Jewish folklore introduced the Golem of Prague, a clay creature animated by linguistic commands, demonstrating an early conceptual desire for natural language prompting.

Karel Čapek, a Czech playwright, introduced the word “robot” in his 1920 play R.U.R., drawing on the Czech word robota, associated with forced labor.

These stories share recurring features that have proven oddly predictive: the artificial being is usually created to serve, frequently exceeds its mandate, and often must be destroyed by its creator. The modern discourse about AI alignment and existential risk is, in many ways, a recapitulation of anxieties that humans have been articulating for at least three thousand years.

By the early 20th century, cinema had completely solidified this mental model. In Fritz Lang’s 1927 masterpiece Metropolis, the “Maschinenmensch“ (machine-human) False Maria established the ultimate, enduring visual language of the humanoid robot. Metropolis was pure science fiction vaporware, but its usability impact was lasting. It conditioned the public to expect that when artificial intelligence finally arrived, it would not be an invisible statistical spreadsheet; it would be an embodied, conversational humanoid. It set the “Anthropomorphic Trap” decades before the first computer existed.

The Mechanical Turk and the Original “Wizard of Oz” Interface

Because the human desire for an intelligent machine was so overwhelming, early entrepreneurs realized they could achieve immense commercial success simply by faking the backend. The most notorious example is the Mechanical Turk, constructed in 1770 by Wolfgang von Kempelen. (Much later, Amazon named a service offering after this proto-AI.)

The Mechanical Turk was presented to the courts of Europe as a magnificent clockwork automaton that could play and win chess against human challengers. It toured the world for decades, baffling audiences and defeating dignitaries like Napoleon Bonaparte and Benjamin Franklin. The reality? It was a brilliantly designed physical UI that hid a massive UX fraud. The cabinet contained a hidden, cramped compartment where a human chess master pulled levers to move the mechanical arms.

From a UX research perspective, the Mechanical Turk was the very first “Wizard of Oz” prototype. In modern UX methodology, a Wizard of Oz test involves a user interacting with what they believe is an automated system, while a human researcher secretly operates the system behind the scenes to test the interaction paradigm before writing expensive code. Von Kempelen weaponized this testing method into a deceptive product. He proved a terrifying psychological law: if the physical interface (the whirring clockwork and the wooden mannequin) projects sufficient complexity, the user will eagerly suspend disbelief and attribute autonomous genius to the machine, entirely ignoring the possibility of a manual backend.

The legacy of the Mechanical Turk is arguably more relevant today than ever. The history of the 21st-century tech boom is littered with “Turk-ware.” Throughout the 2010s and early 2020s, countless startups launched “revolutionary AI” scheduling assistants, receipt scanners, and autonomous grocery checkout systems that were, in reality, just massive offshore teams of human gig workers desperately typing behind a digital curtain.

Babbage, Lovelace, and Genuine Theoretical Proto-AI

Charles Babbage designed the Analytical Engine in the 1830s and 1840s. It was a mechanical, programmable, general-purpose computer, a century ahead of any working equivalent. Babbage never finished building it, partly because Victorian-era machining could not produce parts to the required precision and partly because Babbage was difficult to work with.

Ada, Countess of Lovelace, the mathematically gifted daughter of the poet Lord Byron, wrote extensive notes on the Analytical Engine in 1843. Her notes included what is generally considered the first algorithm intended for execution by a machine, and more importantly for AI history, speculated that such a machine might one day compose music or manipulate symbols of any kind, not merely numbers.

Lovelace also articulated what became known as “Lady Lovelace’s Objection”: that machines could only do what they were explicitly told to do and could not originate anything. Alan Turing engaged directly with this objection in his 1950 paper. The argument continues today, in slightly different form, every time someone debates whether AI is “really creating” or “just predicting.”

Three Lessons from the Proto-AI Era

First, humans wanted artificial intelligence long before they could build it, and that desire shaped what they were eventually willing to accept as the real thing. The cultural appetite for AI predates the technology by millennia. This is part of why ChatGPT was received with such intensity when it launched: it landed in soil that had been prepared by three thousand years of imagination.

Second, the line between real AI and theatrical AI has always been blurry, and audiences have always been willing to overlook the difference. The Mechanical Turk fooled Napoleon. ChatGPT users in 2023 routinely confused fluent text generation with genuine understanding. The deception is sometimes deliberate, sometimes accidental, and almost always commercially advantageous.

Third, our cultural expectations of AI were set by fiction, not by engineering. Researchers must constantly correct misconceptions imported from films like Metropolis, Terminator, and Her. Product strategists should remember this when designing AI interfaces: your users will arrive with expectations shaped by a century of cinema, not by your spec document. You can fight those expectations or design around them, but you cannot ignore them.

The real history of AI begins in 1943. The desire for AI began much earlier, and that desire is part of what we are designing for.

1943: The First Neural Network (in Theory)

The usual starting point for AI history is 1943, when Warren S. McCulloch and Walter H. Pitts published “A Logical Calculus of the Ideas Immanent in Nervous Activity.” McCulloch was an American neurophysiologist and psychiatrist who wanted a mathematical account of the mind; Pitts was a self-taught logician whose early work connected logic, computation, and neuroscience.

Their artificial neuron was not like today’s vast neural networks. It was a simple, idealized model of a biological nerve cell. A neuron could receive inputs, apply a threshold, and fire or not fire. This was important because it showed that networks of simple elements could, in principle, compute logical functions. Thought could be modeled as structured activity in a network. The paper’s own language rested on the “all-or-none” character of neural activity, which made the neuron resemble a logical switch.

The connection to modern AI is direct but incomplete. GPT-class models do not use McCulloch-Pitts neurons in the literal 1943 form. They use numerical units, learned weights, differentiable training, and massive parallel computation. But the founding intuition survived: intelligence may emerge from many simple computational units connected in a network. McCulloch and Pitts gave AI one of its deepest metaphors, and deep learning later turned that metaphor into an industrial system.

By naming their mathematical constructs “neurons,” they birthed the “electronic brain” metaphor, inadvertently inviting humans to anthropomorphize the technology and expect machines to possess human-like common sense.

1949: Hebbian Learning

In 1949, Donald Hebb published The Organization of Behavior. Hebb was a Canadian psychologist and neuropsychologist who sought to explain learning as a biological process in the brain. His central idea is often summarized as: cells that fire together wire together.

Hebb’s claim was that repeated co-activation strengthens connections. This was not a full machine-learning algorithm. It was a theory about learning in nervous systems. But it gave AI a second foundational idea: intelligence should not only be programmed; it should be acquired. A system should improve by changing its internal connections in response to experience.

Modern neural networks are not simple Hebbian systems. A large language model learns through optimization methods that adjust billions or trillions of parameters to reduce prediction error. Yet the conceptual continuity is obvious. Learning means changing connection strengths. What began as a theory of the brain became, after many transformations, a practical method for training artificial systems.

For product strategy, Hebb’s idea matters because it marks the distinction between software that merely executes rules and software that adapts from examples. Most of AI’s commercial power comes from this second category.

1950: The Turing Test and the Usability of Deception

In 1950, Alan Turing published “Computing Machinery and Intelligence.” Turing was a British mathematician who broke the Nazi Enigma code during World War II, and his main research helped define the theory of computation. Rather than asking whether machines can “really” think, he proposed the imitation game: judge the machine by its conversational behavior.

Turing’s contribution was not just philosophical. It was a UX insight before UX existed as a discipline. Users do not inspect a machine’s metaphysical status. They interact with its behavior. They ask whether it answers, helps, misleads, hesitates, recovers, remembers, and adapts. The intelligence of a system is experienced through an interface.

Modern AI has not “solved” Turing’s question in a clean philosophical sense. But ChatGPT made Turing’s framing newly practical. The public did not adopt ChatGPT because it knew how transformers work. People adopted it because it behaved as a conversational partner for writing, coding, planning, tutoring, summarizing, and exploring ideas. The interface made the capability legible.

Turing intuitively understood that natural language is the ultimate user interface. By proposing a text-based dialogue loop as the benchmark, he shifted the definition of AI away from internal architecture and onto user perception, inadvertently inventing the first usability test for AI.

At the same time, Turing established a dangerous UX anti-pattern: human mimicry.

1951: SNARC and Hardware-Software Co-Design

Marvin Minsky was a pioneering cognitive scientist who would shape decades of AI research, and Dean Edmonds was a pragmatically brilliant graduate student who possessed the engineering skills to build physical computing hardware.

Their Stochastic Neural Analog Reinforcement Calculator (SNARC) was the first physical machine built to simulate an artificial neural network, constructed from 3,000 vacuum tubes and surplus airplane autopilot mechanisms to help a virtual “rat” learn to navigate a maze.

SNARC took neural networks out of theoretical mathematics and put them into physical, tangible hardware. Most importantly, it demonstrated reinforcement learning in action; when the virtual rat made a correct turn, the machine adjusted the probability weights of its vacuum tubes, proving a machine could learn via reward mechanisms.

Reinforcement learning is the absolute cornerstone of modern AI alignment and usability. When we use Reinforcement Learning from Human Feedback (RLHF), we are using the software-based descendant of the exact learning mechanism Minsky and Edmonds hardwired into SNARC. The fundamental concept of learning via reward completely carried through.

1952: Arthur Samuel’s Checkers and the Loss of Determinism

Arthur Samuel was an IBM researcher and a pioneer of computer gaming who believed teaching computers to play games was the best path to solving general intelligence.

Samuel wrote a checkers-playing computer program that possessed a scoring function to evaluate board positions and remembered every position it had ever seen. By playing against itself thousands of times, the program learned to play checkers far better than Samuel himself could, prompting him to coin the term “Machine Learning” in 1959.

This was a massive paradigm shift in user capability and mental models. Before Samuel, the heuristic of computing was that a machine could only do exactly what the programmer explicitly commanded. Samuel introduced a new approach: the loss of determinism, proving a machine could independently optimize its behavior.

Self-play reinforcement learning, which Samuel pioneered, is exactly how modern reasoning models discover novel solutions that human engineers never explicitly programmed. Game-playing was a driving force behind many more recent AI breakthroughs, from IBM’s Deep Blue beating the world chess champion in 1997 to Google’s DeepMind doing the same in the more difficult game of Go.

1956: The Dartmouth Conference Names “Artificial Intelligence”

In the summer of 1956, John McCarthy, a young assistant professor at Dartmouth College, organized a six-week workshop with Marvin Minsky (then at Harvard), Claude Shannon (the founder of information theory at Bell Labs), and Nathaniel Rochester (an IBM engineer).

McCarthy coined the term “artificial intelligence” for the workshop’s funding proposal. The conference itself produced no specific breakthroughs, but it established AI as a coherent academic discipline with shared vocabulary and goals.

The Dartmouth attendees were extravagantly optimistic. The proposal described a “2 month, 10 man study of artificial intelligence” and assumed that learning, reasoning, and other aspects of intelligence could be precisely described and simulated by machines. They were off by approximately 70 years, and possibly more. Dartmouth’s importance was institutional as much as technical. It made AI a field. It gave researchers a shared banner, a shared ambition, and eventually shared funding.

Dartmouth also established AI’s recurring weakness: optimism outran capability. The organizers believed rapid progress was possible because they underestimated the complexity of everyday intelligence. For UX and product leaders, this is a permanent warning. A demo can make a capability look solved long before it is reliable enough for ordinary users, hostile inputs, messy data, or high-stakes workflows.

1957: Rosenblatt’s Perceptron

In 1957, Frank Rosenblatt built the first working neural network, the Perceptron. Rosenblatt was a Cornell psychologist and engineer who built one of the first learning machines inspired by neural principles. His 1958 paper described the Perceptron as a probabilistic model for information storage and organization in the brain.

The Perceptron learned to classify inputs by adjusting weights. This was a huge conceptual step. It suggested that machines could learn patterns from data rather than depend entirely on hand-coded instructions. At the same time, early perceptrons were very limited. A single-layer perceptron could only solve a narrow family of problems. Its limitations later became famous.

The modern relationship is clear. The Perceptron was a direct ancestor of neural networks, but not of modern performance. Today’s models are deep, multi-layered, trained on huge datasets, and optimized with sophisticated methods. The early perceptron showed the promise of learned representation. It did not yet have the architecture, algorithms, data, or compute to fulfill that promise.

1958: Lisp and Symbolic AI

While neural ideas were developing, another tradition became dominant: symbolic AI. John McCarthy, fresh from organizing Dartmouth, created the LISP programming language in 1958 as a language for symbolic computation and AI research. The early LISP work at MIT was tied to McCarthy’s “Advice Taker” vision: a program that could represent facts, reason about them, and accept new knowledge.

Symbolic AI treated intelligence as manipulation of explicit symbols. This approach was attractive because it resembled logic and language. Rules could be inspected. Inferences could be traced. Knowledge could be written down. For many years, this was the mainstream of AI research.

Modern AI did not simply discard symbolic AI. It displaced it as the main method for open-ended perception and language, because hand-coded symbolic systems were brittle. They failed when the world refused to fit the tidy categories in the knowledge base. But symbolic methods survive inside modern products as tools, constraints, workflows, schemas, programmatic calls, search indexes, and business rules. A 2026 AI agent using a language model to call APIs is not pure symbolic AI, but it often needs symbolic structure to act reliably.

This is a useful product lesson: neural models are good at interpretation and generation; symbolic systems are good at explicit structure and enforceable operations. Good AI products often combine both.

1966: ELIZA and the First Chatbot Illusion

In 1966, Joseph Weizenbaum introduced ELIZA. Weizenbaum was an MIT computer scientist who later became one of the sharpest critics of AI enthusiasm. ELIZA used pattern matching and scripted transformations to imitate a psychotherapist-like conversation. The program decomposed user inputs and reassembled responses based on rules and keywords.

ELIZA did not understand anything. It used about 200 lines of code. Yet Weizenbaum was disturbed to find that users, including his own secretary, formed emotional attachments to it and confided personal information to it.

ELIZA was technically shallow but experientially powerful. Users attributed understanding to a system that had none. This became known as the ELIZA effect: the tendency to perceive humanlike comprehension in a machine that merely produces plausible conversational behavior.

ELIZA’s method was soon abandoned for serious language understanding. Pattern-matching scripts do not scale to open-ended conversation. Yet ELIZA is one of the most important milestones for UX designers because it showed that interface behavior can create an illusion of intelligence. ChatGPT is not ELIZA; it uses learned language models rather than simple scripts. But every conversational AI product still inherits ELIZA’s danger. Fluency is not the same as understanding, and users will over-trust a well-phrased answer unless the product design helps them calibrate confidence.

1966–1972: Shakey the Robot

At SRI International, Shakey became one of the first mobile robots to combine perception, planning, and action. Charles Rosen, an engineer and founder of SRI’s Artificial Intelligence Center, helped lead the project; Nils Nilsson, an AI researcher and roboticist, contributed to Shakey’s planning methods and later became one of the field’s major chroniclers.

Shakey operated in a simplified world of rooms, boxes, and obstacles. It could perceive parts of its environment, build a model, plan actions, move, and recover from some errors. SRI’s work on Shakey was sponsored by ARPA, later DARPA, and ran roughly from 1966 to 1972.

Shakey mattered because it connected AI to the physical world. It was not enough to solve puzzles on paper. A robot had to sense, decide, and act under uncertainty. Shakey also gave AI planning a concrete domain. The lineage continues into autonomous vehicles, warehouse robots, drones, domestic robots, and embodied agents.

The abandoned part was the assumption that a robot could mostly reason from a clean symbolic model of the world. Real environments are noisy, dynamic, and visually complex. Modern robotics uses learned perception, statistical estimation, simulation, and control methods alongside planning. The dream survived; the early abstraction did not.

1969: Minsky and Papert Kill Neural Networks

In 1969, Marvin Minsky and Seymour Papert published Perceptrons. Papert was a South African-born mathematician and computer scientist who co-founded MIT’s AI Lab and later became famous for Logo and constructionist learning. Their book rigorously analyzed what single-layer perceptrons (like Rosenblatt’s pioneering 1957 work) could and could not do.

The book is often blamed for chilling neural-network research. That is partly unfair and partly true. Minsky and Papert did not prove that all neural networks were useless. They showed that the perceptrons of the time had severe limitations. But because multilayer learning methods were not yet practical, the field shifted toward symbolic AI.

In retrospect, this episode is a lesson in timing. A critique can be correct for the systems available at the time and still not predict the future. Neural networks needed better algorithms, more data, and far more compute. Once backpropagation and GPUs arrived, the old limitation no longer defined the field.

The Perceptrons book is a cautionary tale about how a partial truth, presented authoritatively by famous researchers, can derail an entire field for a generation. A technology may fail in one decade because the surrounding stack is not ready, then become dominant decades later.

1972: SHRDLU

Terry Winograd, then a graduate student at MIT, built SHRDLU as his doctoral thesis. It was a natural language program that operated in a simulated blocks world of geometric shapes.

SHRDLU could understand and execute commands like “Put the red block on top of the green pyramid.” It maintained a world model and could answer questions about its actions. To observers in 1972, it seemed like a major step toward genuine language understanding.

In retrospect, SHRDLU was a dead end. Its capabilities did not generalize beyond the blocks world. The hand-coded knowledge structure that made it work could not scale to real-world language.

The lesson echoes through to 2026: systems that work in narrow demonstration domains often fail to scale. Modern language models work because they were trained on a substantial fraction of all human-written text, not because anyone hand-coded their knowledge. Winograd later became one of the founding figures of HCI at Stanford and was Google co-founder Larry Page’s PhD advisor.

1972–1973: Prolog, Logic Programming, and the Lighthill Report

In the early 1970s, logic programming became a major AI idea. Alain Colmerauer, a French computer scientist, and Robert Kowalski, a logician and AI researcher, helped establish Prolog and the view that computation could be expressed as logical relations. Logic programming later influenced European AI and the Japanese Fifth Generation project.

Today, Prolog is a footnote. Modern AI does not work by formal logical inference. But Prolog’s legacy persists where formal reasoning is genuinely required: theorem provers, certain database systems, and specialized regulatory and verification applications.

Mid-1970s: First AI Winter

At almost the same time, AI suffered a public setback. In the United Kingdom, Sir James Lighthill, a British applied mathematician, produced a critical review of AI research. The Lighthill Report, submitted in 1972 and published in 1973, criticized the field’s progress in areas such as robotics and language processing and helped trigger reduced support for AI research in Britain.

DARPA funding in the United States was cut sharply. Many AI researchers reframed their work as something else. This was the first “AI winter.” It lasted roughly from 1974 to 1980.

The contrast is revealing. AI had beautiful formalisms, but messy reality resisted them. Systems worked in small worlds and failed in broad contexts. This gap between demo and deployment would recur for decades.

Modern AI is vulnerable to the same pattern. Benchmarks and demos are useful, but real products face distribution shift, user misunderstanding, malicious use, latency, cost, privacy, integration, and accountability. Lighthill’s critique was harsh, but the underlying product question was valid: where was the reliable value?

Late 1970s-1980s: Expert Systems

Edward Shortliffe, a Stanford physician-researcher, built MYCIN in the early 1970s as part of his PhD work. It was an expert system for diagnosing bacterial infections that incorporated about 600 rules captured from interviews with infectious-disease specialists.

MYCIN advised physicians on bacterial infections and antibiotic treatments. It used a knowledge base of rules and an inference engine, along with certainty factors to express imperfect confidence. It could also explain aspects of its reasoning, which made it unusually relevant to high-stakes decision support.

MYCIN performed approximately as well as human experts in controlled tests. It was never deployed clinically, partly because of liability concerns about who would be responsible if its recommendations harmed a patient. (Forty years later, this exact question still complicates clinical AI deployment.)

Expert systems became the dominant form of commercial AI in the 1980s.

The expert systems approach of encoding human expertise as explicit rules was eventually abandoned because it did not scale. Capturing knowledge from experts was expensive, the rules were brittle, and the systems could not handle situations their authors had not anticipated.

But the expert systems era left two important legacies. First, it established the commercial viability of AI, paving the way for later investment. Second, it taught the field, painfully, that explicit rules cannot capture the full subtlety of human expertise. This lesson directly motivated the shift to learned systems that now dominate AI.

1980: XCON and the First Big Commercial AI Payoff

Around 1980, Digital Equipment Corporation began using XCON, also known as R1, to configure VAX computer systems. John McDermott, a Carnegie Mellon AI researcher, was the central figure behind R1/XCON. The system translated customer orders into valid computer configurations, catching missing or inconsistent components before manufacturing and delivery.

XCON was the first AI system to achieve massive, undeniable commercial success, saving DEC an estimated $40 million a year. It succeeded where MYCIN failed because of its bounded context; the domain of DEC computer parts was finite, knowable, and heavily structured, preventing the catastrophic edge-case failures seen in broader systems.

It did not need to understand the bigger world. It needed to solve a painful business problem: configuring complex products correctly. This is one of the most important product lessons in AI history. Narrow, expensive, repetitive, expertise-heavy tasks are excellent AI targets.

XCON also demonstrated the limits of rule-based AI. As products changed, rules had to be updated. Maintenance became expensive. The system’s value was real, but the approach did not generalize easily. In modern terms, XCON was a successful vertical AI product with high operational burden.

1982: Japan’s Fifth Generation Computer Project