UX Roundup: GPT-6.2 | Prepare with AI | Speed | Unambitious Academics | AI Permission Requests | Permissions Barrage | Recognition-Primed Decisions | Directed Editing | Misleading Defaults

Summary: GPT-6.2 Astra underwhelms at agentic video production | Use AI to prepare for human conversations | Speed is good UX | Academics aim AI at old chores | Users rubber-stamp AI permission requests | Ask for permissions in context, one at a time | Experts decide at a glance, and so do AI users | Sliders suggest edits users want but wouldn’t request | Never precheck paid add-ons

UX Roundup for November 20, 2026 (GPT Image 2.6)
GPT-6.2 Astra: Not Impressed
OpenAI has released its strongest model to date: GPT-6.2 Astra. It’s certainly an upgrade. But I am not impressed, because it still sits a rung below its main competitor, Anthropic’s Fable 5.2.

An upgrade can be real and still leave the reviewer unmoved. (GPT Image 2.6)

The version number is 6.2. So, by coincidence, is the mood at the judges’ table. (GPT Image 2.6)

A bigger version number doesn’t lift a model onto a higher ledge. (GPT Image 2.6)
You be the judge. Compare the explainer video Service as a Software (YouTube, 3 min.), which I made with GPT-6.2 Astra, with the training video Closed vs. Open Questions in User Research (YouTube, 24 min.), which I made with Fable 5.2, or even with Mental Models, Conventions, and Expectations in UX Design (YouTube, 6 min.), which I made with Anthropic’s “discount” model, Opus 5.5.
Viewing statistics show that people watched Mental Models, Conventions, and Expectations in UX Design 47% longer on LinkedIn than they watched Service as a Software. On YouTube, Mental Models wins with a whopping 297% longer view time. It’s possible that mental models are more thrilling than the future of software, but I mainly credit the higher audience engagement to better agentic editing by Opus 5.5.

Compare the videos for yourself. Eyes open, please. (GPT Image 2.6)

Even Anthropic’s “discount” model made a better video than OpenAI’s brand-new flagship. (GPT Image 2.6)
OpenAI’s launch benchmarks beg to differ. On Agents’ Last Exam, where agents carry out multistep professional work, the company claims a 5.5-point lead for Astra over Fable 5.2, at only 5% of the cost.
Such benchmarks increasingly hand out trophies for solving theoretical puzzles, while the tasks real users face go ungraded. My sample covers only one video per model, but I’ll take work I can watch over a leaderboard any day.

OpenAI’s benchmarks crown Astra the champion. My monitor disagrees. (GPT Image 2.6)

Benchmarks grade the theory exam. Real users need a model that can swim. (GPT Image 2.6)
Call me an ungrateful bastard: motion graphics generated entirely by a language model still leave me grumbling about quality. It’s like griping about the legroom on a flying carpet. Just months ago, such animations required painstaking human control. Now AI does that work for us, which is a small miracle.

Complaining about AI-made motion graphics is like griping about the legroom on a flying carpet that serves a nice cuppa. I do it anyway. (GPT Image 2.6)

The language model now directs the motion graphics. The human’s remaining job is to complain. (GPT Image 2.6)

AI video production is a miracle. That doesn’t stop me from ranking the miracles. (GPT Image 2.6)
For now, Anthropic’s months-old models produce better videos than OpenAI’s most advanced model, released this week. Do better, Sam!

The newest and flashiest model isn’t automatically the best at the job at hand. (GPT Image 2.6)

An upgrade doesn’t necessarily put a model at the top of the class. Do better, Sam! (GPT Image 2.6)
Tokenaires: OpenAI’s Heaviest Users Each Run $23,000 of AI per Day
In September, I reported that OpenAI’s 90th-percentile researcher ran through $7,047 worth of AI per day at list prices. That now looks quaint. OpenAI’s new report puts the figure at an eye-popping $23,000 for October, up 105% from September’s $11,000. (September’s figure was depressed by a temporary efficiency tweak, so discount the doubling a bit. Still, usage has more than tripled since my mid-August number.)

At list prices, OpenAI’s heaviest AI users each run $23,000 of AI per day. Finance chiefs elsewhere may need to sit down. (GPT Image 2.6)

Usage that doubles every month follows the old legend of the chessboard: the first squares look harmless, and then the towers leave the picture. (GPT Image 2.6)

A temporary efficiency tweak depressed OpenAI’s token spending in late summer. It was rolled back (for unrelated reasons), and the climb resumed. (GPT Image 2.6)
Is that much AI worth it? In a survey of 101 OpenAI research staff members, the median respondent (the typical researcher, not the heaviest user) would give up 40% of the team’s headcount to keep frontier coding agents for 3 months. Hypothetical and self-reported, but telling.

The median researcher in OpenAI’s survey would give up 40% of the team to keep frontier coding agents. The rest of the formation reports no complaints. (GPT Image 2.6)

The 40% figure is a median: half of the surveyed researchers would give up at least that much of their team. (GPT Image 2.6)

The trade is hypothetical and self-reported, which is a caveat for readers and small comfort for Kevin. (GPT Image 2.6)
Few companies will pay $23,000 per employee per day. The top labs don’t, either: they get tokens below wholesale, and OpenAI’s researchers have no usage quotas. This is the two-tier AI world I predicted for 2026, now dividing companies as well as subscribers. The labs’ tokenaires sit on top with the taps wide open. Everybody else must compete on a trickle, a punishing handicap.

The two-tier AI world, hive edition: OpenAI’s researchers have no usage quotas, while everybody else lines up for a trickle. (GPT Image 2.6)

Outside the top labs, tokens get rationed. Few companies will pay $23,000 per employee per day. (GPT Image 2.6)
Even if usage stops doubling monthly, I expect the tokenaires to reach $100,000 per person per day in early 2027, or mid-2027 at the latest. (The slow scenario needs only 20% monthly growth.)

In September, I called OpenAI’s spending data a weather forecast for your power users. The updated outlook: $100,000 per person per day by early 2027. (GPT Image 2.6)

Even if growth slows to 20% a month, about a fifth of October’s pace, the heaviest users reach $100,000 per day by mid-2027. (GPT Image 2.6)
Raise your AI budget per employee now. The upper tier isn’t waiting.

Asking for an annual AI budget equal to what a top OpenAI researcher burns through in one day can still alarm a board. Ask anyway. (GPT Image 2.6)
Another data nugget from the OpenAI report: autonomous success on research tasks requiring two to four human hours rose from 23% in March to 46% in October. In only 7 months, AI got twice as good at these medium-complex tasks. You do get something for your money with AI. Can we reach week-long tasks (40 human hours) by year-end as I predicted in January? Probably not, but given AI’s improvement rate, this milestone can’t be far away.

Raise your AI budget per employee now. My guess: the best candidates will soon ask about tokens before titles, because tokens will be the main way to get good work done and thus the main way to advance in your career. (GPT Image 2.6)
Talk to AI Before You Talk to Humans
A short prepversation with AI improves the human conversation that follows. Two new randomized studies from different Google labs show this for doctor visits and for group discussions of contested policies. In both, the winning AI was the better questioner.

TV guests have always had producers who brief them in the green room on what the host will ask. AI gives everybody a pocket-sized green room. (GPT Image 2.6)
Doctor Visits: Same Model + Better Script = 17 Points More Prepared
US primary-care visits typically last 15–20 minutes, and patients often blank on what they meant to ask. Spend cheap AI minutes to make expensive human minutes count. Many patients already try. According to Gallup, 59% of Americans who recently used AI to research health information did so before doctor visits.

A 15-minute visit can’t absorb an endless scroll of worries. The visit-prep agent’s instructions told it to identify the patient’s 3 most important questions. (GPT Image 2.6)

Even people with excellent memories go blank in the exam room. A written prep sheet works as external memory, for fish and humans alike. (GPT Image 2.6)
In a randomized study by Xincheng Tan and colleagues at Google Research, 802 US adults prepared for doctor visits in the Google Health app with either a minimally prompted chatbot (roughly what general-purpose AI gives people today) or a structured visit-preparation agent. Both versions ran on Gemini 2.5 Flash with a running prep sheet; only the system prompt differed. The agent’s script guided it to interview patients, explain what would likely happen at the visit (a mental model of the encounter), and build a prioritized list of questions for the clinician.

Clinicians call it the doorknob question: the real concern a patient raises only as the doctor reaches for the door. A prioritized question list moves it to the start of the visit. (GPT Image 2.6)

Nassim Taleb’s famous turkey grew more confident with every meal, right up to the Wednesday before Thanksgiving. Accurate expectations are the most underrated part of preparation. (GPT Image 2.6)
Result: 74% of agent users felt prepared, vs. 57% with the chatbot, and blinded clinicians rated 67% of agent sessions comprehensive, vs. 51%, with equally high safety. Thus, a stronger script coaxed more from the same model, which already looks primitive by today’s standards.

Once patients know what will happen at the visit, the routine parts go faster. Sometimes the doctor becomes the least-prepared person in the room. (GPT Image 2.6)
Group Discussions: A Thinking Partner Beats a Worksheet
Haiwen Li and co-authors from Google DeepMind and MIT placed 1,173 Americans in 391 three-person text chats about contested policies.
Before each 10-minute discussion, participants spent 7 minutes in one of three conditions: watching an irrelevant video (the control group), answering three reflection questions in writing, or discussing those questions with an AI reflector. The researchers instructed the reflector to remain neutral, ask how the policy would work in practice, and invite users to imagine the other side’s reasoning.

The control group watched an irrelevant 7-minute video, so every condition took the same amount of time before the discussion. Somebody has to be the guinea pig. (GPT Image 2.6)

Asked to imagine the other side’s reasoning, AI-reflection participants reported understanding their opponents’ reasons better than did people who reflected alone or watched a video. (GPT Image 2.6)

Groups whose members agreed from the start became more extreme during the discussion, whatever their preparation. Deliberation needs a dissenter. (GPT Image 2.6)
Only AI reflection improved the discussions: 62% of AI-prepped groups beat the median video group on discourse quality.
Why did the worksheet fail? Solo reflection mostly rehearses your own case, and it even nudged stances slightly further toward the extremes. AI also prompted people to reconsider sooner: 30% of its users shifted their stance during reflection, vs. 17% of video watchers, though the total change by the end was about the same. Opinions then converged most in the AI-prepped groups.

Reflecting alone mostly rehearses your own case: the written-reflection group ended up with slightly more extreme stances and no better discussions than the video group. (GPT Image 2.6)

AI reflection prompted earlier changes of mind: 30% of AI users shifted their stance before the group even met, vs. 17% of video watchers. (GPT Image 2.6)
(Socrates called his method midwifery, because he helped people deliver their own ideas. Now the midwife fits in a phone.)

Socrates asked questions until people delivered their own ideas. Athens thanked him with a death sentence; today’s participants rated the AI reflector as helpful. (GPT Image 2.6)

The AI reflector was told never to praise or validate users. Most participants still said it helped them think more carefully. Sycophancy is optional. (GPT Image 2.6)
How to Design a Prepversation
The two studies agree on the recipe:
Script the flow. In the medical study, the script was the entire difference.

A better script squeezed more out of the same model: 74% of patients felt prepared with the structured agent, vs. 57% with the minimally prompted chatbot. (GPT Image 2.6)
Ask more than you tell. The visit agent helped patients formulate their own questions for the doctor, and the reflector even declined to answer factual questions and sent users off to search for themselves: 10% looked things up, vs. 4% of the video group.

The AI reflector asked people how a policy would work in practice, step by step. Explaining the mechanism punctures the illusion of explanatory depth. (GPT Image 2.6)

Asked for facts, the reflector invited people to search on their own: 10% of its users looked things up, vs. 4% of the video group. (GPT Image 2.6)
Keep it short. A concise agent aiming for under 70 words per turn did about as well as the wordier standard version.

The DeepMind researchers suggest a 30-second reflection prompt before people post online comments. Imagine the comment sections. (GPT Image 2.6)
Measure what happens between people. Patient satisfaction didn’t correlate with clinicians’ quality ratings, and neither study measured health outcomes or decision quality. Next study, please.

Patients’ satisfaction ratings didn’t correlate with clinicians’ quality ratings of the same sessions. Measure both. (GPT Image 2.6)
Run a prepversation before your next important meeting. Tell your AI, “Interview me, one question at a time, to prepare me for this meeting. Then give me a one-page summary.” Rehearse with inexpensive AI before spending an expensive human hour.

Run a prepversation before any high-stakes human encounter, including the holiday dinner table. (GPT Image 2.6)
Speed Is a Feature
Before anyone admires your typography or reads your value proposition, every visitor experiences the same opening act: the wait. Speed creates the first impression and renews it on every page view.
A fast site feels competent and trustworthy; a slow one feels broken. And few people will type a credit card number into a site that feels broken.
Slow pages bleed visitors. Google reported in 2016 that 53% of mobile visits get abandoned when a page takes longer than 3 seconds to load. Hardware has kept improving since, but pages squander the gains on heavier scripts, fonts, and trackers.
Treat speed as a core design requirement. Set a performance budget and give it veto power over features, because every image, font, and third-party script carries a price tag in milliseconds. Then test on a mid-range phone over a cellular connection, where your customers live.

Every loading second gives the customer another second to reconsider. (GPT Image 2)
Scientists Still Use AI as a Spyglass: Only 0.4% of Requests Ask It to Do the Science
OpenAI’s logs from 1,819 scientists show AI mostly explaining concepts and drafting papers. A third vendor now confirms what Google and Anthropic found: ideation is scarce, and checking AI’s output is the top complaint. New twist: when scientists brainstorm, they usually let AI pick the question.
In August 1609, Galileo pitched his telescope to the Venetian Senate as a spyglass for spotting enemy ships. Scientists aim AI just as low today, pointing it at old chores they want done faster.

The telescope began its career as a naval spyglass. Patrons always want the old job done faster. (GPT Image 2.6)
Chris Liao and colleagues at OpenAI classified 580,186 scientific messages (76-page PDF) that 1,819 academic scientists sent over 5 weeks in 2026 and surveyed 251 of the scientists. (Caveat: they signed up for free frontier-model access, so they skew toward enthusiasts.)
Reading and writing account for more than half of usage, and the top tasks were explaining scientific concepts (19% of messages) and drafting manuscripts (14%). Only 0.4% of messages asked AI to run an entire research project, though 15% of the scientists tried it at least once.

Explaining scientific concepts is the top task, at 19% of scientists’ messages. Even veterans of the field want a tutor. (GPT Image 2.6)

Requests for AI to run an entire research project make up 0.4% of scientists’ messages. The rest of the time, the miracle fixes the bibliography. (GPT Image 2.6)
Third Vendor, Same Verdict
My analysis of AI-accelerated science earlier this week relied on Google and Anthropic data. OpenAI’s numbers rhyme:


Scientists follow up to clarify, check, or revise after 32% of their requests. Skepticism is one research task nobody delegates. (GPT Image 2.6)

AI made hypotheses cheap, but 44% of the surveyed scientists say funding limits their progress, and 52% name time or staffing. (GPT Image 2.6)
The checking row is a design brief: show sources and working steps, so verification takes seconds instead of hours.

Time spent checking or correcting AI’s output is scientists’ top complaint, cited by 45%. Every lab still needs a coal-mine canary. (GPT Image 2.6)
But one comparison breaks the pattern. Reading makes up 28% of ChatGPT’s science messages, compared with 7.6% of Claude’s research sessions, where analysis and writing dominate. Perhaps ChatGPT got typecast as the tutor.
Usage logs measure the tool as much as the user, and your own product analytics have the same blind spot. So trust the findings that survive across vendors.
Curiosourcing = Letting AI Pick the Question
Who steers what little ideation occurs? In brainstorming, 69% of messages ask AI to choose the research question instead of supplying one. Such curiosourcing is risky with current tools, because 49% of the scientists in Google’s survey said AI steers them toward safer questions (28% said riskier).
Don’t settle for one suggestion. Demand a ranked slate of 10 questions, each labeled with its risk, potential payoff, and the experiment that would kill it. Tool designers: make that the default.

Curiosourcing: in brainstorming, 69% of scientists’ messages ask AI to choose the research question. (GPT Image 2.6)

In Google’s survey, 49% of scientists said AI steers them toward safer questions. Curiosourcing rarely takes the Moonshot exit. (GPT Image 2.6)

Don’t settle for one suggested question. Demand a ranked slate that shows each idea’s risk, potential payoff, and the experiment that would kill it. (GPT Image 2.6)
Conclusion: Point It at the Sky
Of the surveyed scientists, 98% expect AI to accelerate discovery, yet their ambitions stay narrow, with faster task completion accounting for 50% of their stated reasons and automated research loops for only 2.2%. Optimism shrinks where the labjam grows: 31% of mathematicians and statisticians expect a 10-year discovery within 3 years, vs. 10% of engineers and materials scientists, whose atoms stay expensive.

When 98% of surveyed scientists agree that AI will accelerate discovery, take note: academics rarely agree this much about anything. (GPT Image 2.6)

Among mathematicians and statisticians, 31% expect a 10-year discovery within 3 years, compared with 10% of engineers and materials scientists. Proofs fly, but dams must hold water. (GPT Image 2.6)

The labjam: ideas arrive at machine speed, but experiments still run one bench at a time. (GPT Image 2.6)
That’s spyglass thinking. But less than 5 months after Venice, Galileo aimed his instrument at Jupiter and found 4 moons. Since I expect superintelligence by 2030, that 0.4% won’t stay small for long.

Requests for AI to run a whole study are only 0.4% of scientists’ messages today. Don’t judge a technology by its caterpillar stage. (GPT Image 2.6)
21% of Developers Admit to Rubber-Stamping AI Agents. More Do It.
When asked, 21% of software developers admit to approving AI agents’ permission requests by default. Believe the confession. Doubt the count: the true share of reflex approvers is guaranteed to be higher.

21% of developers admit to approving AI agents’ requests by default. Every confession is real; the count is a floor. (GPT Image 2.5)
Larissa Salerno and co-authors from Chalmers, the University of Melbourne, and Eindhoven University of Technology interviewed 18 developers who use coding agents such as Claude Code, then surveyed 115 practitioners. Only 48% said they read permission requests carefully, while 34% skim them and 34% have grown less careful over time. After a run of look-alike requests, one interviewee “stopped caring.”

Only 48% of developers say they read permission requests carefully. That’s the self-reported number, so the real number is probably lower. (GPT Image 2.5)

34% of developers skim permission requests, and 21% approve them by default. Speed gets rewarded; reading doesn’t. (GPT Image 2.5)
Admitted < Actual
Why am I so sure? Three biases suppress the count. Few people confess to a sloppy habit they don’t have, while many hide one they do:
Embarrassment. Waving through security prompts sounds negligent, so some reflex approvers keep quiet.
Amnesia. Approving on allowpilot is automatic and leaves little trace in memory. You can’t report decisions you never consciously made.

Allowpilot approvals happen below awareness, so even honest survey respondents can’t report them. (GPT Image 2.5)
Self-selection. A permissions survey attracts the developers who care most about the topic, a skew the authors acknowledge.

A survey about permission habits attracts respondents who care about permissions, so its numbers flatter everybody else. (GPT Image 2.5)
These biases leave the study useful and indicate which way to correct it. So treat the 21% as a floor. (Even the authors think their 34% carelessness figure is too low.) Self-assessment is just as rosy on comprehension: 73% of survey respondents claimed full understanding of their agents, but almost half the interviewees, probed in depth, admitted partial understanding or none.

73% of developers claimed full understanding of their AI agents. Probed in depth, almost half the interviewees admitted partial understanding or none. (GPT Image 2.5)
Behavioral data tells a franker story: Anthropic reported in March that Claude Code users approve 93% of permission prompts. When almost every request gets a yes, yes becomes a reflex.

When almost every request gets a yes, saying yes becomes a reflex, and reflexes don’t stay at the office. (GPT Image 2.5)
Pavlov would recognize the conditioning, and my old slogan applies: trust what users do, not what they say.

Trust what users do, not what they say. Self-reports flatter; behavior testifies. (GPT Image 2.5)
Fewer Prompts = Real Oversight
Pep talks won’t fix this. As I argued in “Watch Duty & Verdict Duty,” vigilance decays with every routine approval, so tier requests by consequence and pass only the important ones to a human. My analysis of the Sidekick study reached the same verdict for agent monitoring: interrupt only for decisions.

Tier requests by consequence and reserve human review for the important ones, leaving enough attention to judge them well. (GPT Image 2.5)
The survey data maps onto three tiers:
Allow silently: read-only actions (98% readily grant them) and work inside a sandbox. Stop asking.

Read-only actions and work inside a sandbox can proceed without asking. Just make sure the sandbox has walls. (GPT Image 2.5)
Ask, and show the consequences: writes and deletions (which 77% already scrutinize more closely) plus anything off-task. Say what will change and whether it can be undone.

When an agent must ask, it should summarize what will change and whether it can be undone; users don’t need a transcript of its reasoning. (GPT Image 2.5)
Block by default: sensitive data (62% would never grant access) and root access (56%). And make a “no” cover the goal, so the agent can’t detour around it.

Some requests deserve no prompt at all: 56% of developers would never grant root access, so block it by default. (GPT Image 2.5)

A “no” must cover the user’s goal as well as the immediate command, or the agent will find another route. (GPT Image 2.5)
Finally, as the authors warn, don’t let adaptive permission systems learn from the rubber stamp. An allowpilot click is fatigue posing as preference.

Every needless “Allow?” trains users to approve risky requests without thinking. (GPT Image 2.5)
Request Permissions in Context, One at a Time
Permission overload predates AI agents, which are now making it worse. An app that fires 6 permission prompts at first launch is asking a stranger for the house keys before saying hello. Users tap Deny, or they tap Allow just to make the popcorn stop. Neither choice amounts to informed consent. Adrienne Porter Felt and colleagues at UC Berkeley found in 2012 that only 17% of Android users paid attention to permission requests, and only 3% could answer all three basic questions about what the permissions meant. Prompts fired out of context become noise, and noise gets swatted.
On iOS, the system asks once; deny the request, and the app must send the user digging through Settings to reverse the decision.
Spend each prompt where it counts. Request the camera when the user taps the camera button and location when the user asks for directions. Explain the benefit in one sentence, and never show more than one prompt per task. Ask when the user wants the feature; anything earlier is harassment with a dialog box.

Allow, allow, deny, allow: the rhythm of a reflex. AI or not. (GPT Image 2)
15 of 16 Users Wanted AI Edits They’d Never Think to Ask For
You know when an AI draft is off. But naming what’s off is harder. A chat box makes you recall the fix and spell it out, violating my 1994 usability heuristic of recognition rather than recall. (The authors cite it, so for once I’m not the one bringing it up.)

Users can’t describe the tone they want, but they’ll pick it out of a lineup in a second. Recognition beats recall. (GPT Image 2.5)
Ajit Mallavarapu from Cornell Tech and Ziwei Gu from Harvard built DimSteer. It samples 30 drafts per prompt and identifies the 1–3 directions in which the model’s internal activations vary most. It then labels these directions (e.g., Formality, Urgency, or Dissonance) and turns them into sliders that steer the model directly, without reprompting.

Machine-named sliders are hypotheses, and some labels explain more than others. (GPT Image 2.5)

Each prompt gets its own sliders: for an urgent outage message, the model surfaced Urgency, Formality, and Directness. (GPT Image 2.5)

Revising by prompt piles up a scroll of rephrased wishes. Revising by slider takes a drag. (GPT Image 2.5)
In tests with 16 frequent AI users, sliders beat a matched chat interface, cutting median mental demand from 5 to 2 on a 7-point scale and time per task by 44%, with no loss in perceived success.
All 16 said the sliders offered changes they wanted, yet 15 of 16 wouldn’t have thought to ask for them in a prompt. One user said of the Dissonance slider: “I couldn’t think of how to describe it… but I like that slider.”

Sliders cut the time per task by 44%, and they work for any arm length. (GPT Image 2.5)

All 16 participants wanted the changes the sliders offered, yet 15 of 16 wouldn’t have thought to ask for them. (GPT Image 2.5)
The empty prompt box induces promptnesia: users can’t recall the words for what they want, though they recognize it on sight.

Promptnesia: facing an empty prompt box, users can’t recall the words for what they want, so they type “make it better.” (GPT Image 2.5)
Caveats: a mostly young sample; short texts; self-rated success; sliders that sometimes changed more than their labels promised; and a sluggish prototype that needed 15 seconds per slider move, nearly 150 times the 0.1-second limit for direct manipulation. I predict that real-time steering will widen the gap.

Each slider move took 15 seconds, nearly 150 times the 0.1-second response that makes direct manipulation feel direct. (GPT Image 2.5)
Users split 8 to 8 on which interface they’d pick for their own writing, and several asked for both. That’s the hybrid UI I predicted in my article on the articulation barrier. Chat revived the DOS prompt; sliders revive the GUI. Your AI product needs both.

Chat brought the command line back from near extinction. Sliders bring back the GUI. (GPT Image 2.5)

Automatically discovered sliders won’t cover every wish: 5 of 16 participants defined a custom dimension. Keep a way to say what the knobs can’t. (GPT Image 2.5)
Show the knobs, keep the box.
Never Precheck a Box That Costs Money
Defaults dominate: most users accept whatever the interface preselects, as I documented in my Defaults article. That power makes prechecked add-ons both profitable and poisonous.
A default is a recommendation users trust you to make in their interest. A prechecked $19.99 warranty, a $4.99 monthly text-alert subscription, and a “share my data” box recommend something in yours instead. The user who spots them feels ambushed; the one who doesn’t feels robbed when the statement arrives. Both stop trusting the checkout, and a distrusted checkout gets audited screen by screen.
The European Union settled the legal question in 2014, when Article 22 of the Consumer Rights Directive took effect and entitled consumers to a refund of any extra payment imposed through a default they had to reject.
Even without regulators, the arithmetic works against you, because the add-on revenue from inattentive buyers is small change next to the losses from abandoned carts and chargebacks. Anything that costs money starts unchecked. Sell the warranty with an honest pitch, or don’t sell it.

Total: your items, plus everything you didn’t notice. (GPT Image 2)
Recognition-Primed Decisions: Experts Don’t Compare Options, and AI Users Approve at a Glance
Experts under time pressure skip the comparison shopping. They recognize the situation, imagine the first plausible action, and act. Good UX feeds that recognition with stable, glanceable cues. With AI, a wrong glance costs more, because users vet AI the same way, and people who approve 97% of an agent’s requests rarely catch the dangerous one.

Experts under time pressure don’t hold a beauty contest among options. While the analysis gets organized, the veteran has already recognized the fire and put it out. (GPT Image 2.6)
Definition: The recognition-primed decision (RPD) model describes how experienced people decide under time pressure. Experts recognize the situation as a familiar type; the pattern suggests a typical action, which they mentally simulate. If the simulation passes, they act. If it reveals a snag, they adjust or try the next candidate. Experts screen options one at a time, and the first workable one wins.

The expert tests a single option by imagining it. When the bridge snaps in the simulation, she moves on to the next candidate without ever drawing up a comparison matrix. (GPT Image 2.6)
This is satisficing with a built-in quality check, and it clashes with the business-school ideal: list the alternatives, weight the criteria, compute the optimum. Real experts skip nearly all of that. Recognition carries the expertise.

The business-school ideal says to list every alternative, weight every criterion, and compute the optimum. Reality tends to arrive before the spreadsheet is done. (GPT Image 2.6)
Deciding Is Seeing
The model was born on the fireground. In 1985, Gary Klein and colleagues interviewed 26 fire commanders averaging 23 years of experience, probing 156 tough decision points.
Klein expected commanders to weigh at least two options. Instead, they compared options in fewer than 12% of cases; in more than 80%, they recognized the fire as a familiar type and ran its standard response.
The commanders denied making decisions at all. They looked, they knew, they acted. Chess masters do likewise: Adriaan de Groot showed in 1946 that they out-recognize weaker players without out-calculating them.

Klein’s fire commanders denied making decisions at all. With 23 years of patterns in their heads, they looked, they knew, they acted. (GPT Image 2.6)
The lesson for designers: experts decide while assessing the situation. By the time a hand reaches for a control, the choosing is over. Yet most design effort goes into the action side (buttons, menus, flows), while the bigger payoff sits on the situation side, in what a screen lets a practiced eye recognize.

Design effort pours into the action side of the interface: buttons, menus, flows. But the expert’s decision is made on the situation side, which too often gets one foggy mirror. (GPT Image 2.6)
Intuition Is Earned From Stable Cues
Such intuition rests on experience. In 2009, Daniel Kahneman, the great skeptic of expert judgment, and Klein published a joint verdict: recognition deserves trust only when the environment offers stable, valid cues and the person has had long practice with clear feedback. Firefighting and chess qualify. Stock picking doesn’t.

Kahneman and Klein agreed that intuition deserves trust only where cues are valid and feedback is clear, which admits chess and excludes stock picking. (GPT Image 2.6)
A consistent UI provides that environment, because every stable screen deposits recognition capital: a growing library of patterns that experienced users can match to new situations, with the balance compounding through repeated encounters. The nurse on her 4,000th shift recognizes your dashboard before she could read a single label. RPD stretches my old heuristic of recognition rather than recall from remembering commands to judging entire situations.

Every stable screen deposits recognition capital. After 4,000 shifts, the nurse recognizes the one thing that’s wrong without reading the other 11. (GPT Image 2.6)
Don’t neglect the model’s second half, in which experts test the primed action by imagining it. Make that imagining cheap.

Experts evaluate the primed action by imagining it. A preview makes that imagining cheap, and far cheaper than a flooded castle. (GPT Image 2.6)
Recognition Fails Fast and Confidently
Three failure modes haunt shipping products:
• No patterns. Novices have no prototypes, so the dense dashboard that a veteran reads in one glance is noise to them.

A novice owns no patterns, so the dense display that a veteran reads in one glance is noise to a newcomer. Scaffold beginners separately. (GPT Image 2.6)
• Burned patterns. A redesign demotes every expert to novice overnight. Worse, the old patterns still fire, and a misleading pattern beats no pattern for speed while losing badly on accuracy.

A redesign demotes every expert to novice overnight. Worse, the old patterns still fire, and now they point at the moat. (GPT Image 2.6)
• Counterfeit patterns. When a rare, dangerous state resembles a routine one, the expert runs the routine playbook. If “delete one file” and “delete entire project” share a dialog template, recognition will autocomplete the disaster.

When a dangerous action follows the same template as a routine one, recognition will autocomplete the disaster. Make different things look different. (GPT Image 2.6)
AI Applications: Polish Is Now a Counterfeit Cue
With AI, the machine proposes and the human disposes. That hands users the last step of RPD (evaluate one candidate) while skipping the first (size up the situation). But nobody can mentally simulate an action in a situation he or she never assessed.
Users then fall back on glanceptance: output that looks like a typical good answer gets accepted at a glance. Polish used to be a fairly valid cue, since fluent prose and tidy formatting took competence to produce. But AI makes polish free.

Glanceptance: output that looks like a typical good answer gets accepted at a glance. AI made polish free, so polish no longer proves anything. (GPT Image 2.6)
The remedy is to give the practiced eye something valid to recognize. Have the AI state its situation assessment before its answer (“treating this as a routine refund”), because an expert spots a wrong frame faster than a wrong detail.
For novices, that statement is a loan from a pattern library they don’t yet own. Fixed positions let readers recognize the verdict, evidence, and caveats at a glance; a fresh wall of prose forces them to start reading from scratch.
Generative UI that redraws the screen for every request banks nothing. Users also bank patterns about the AI itself, so a silent model update burns that capital as surely as a redesign.

Have the AI state its situation assessment before its answer. An expert spots a wrong frame faster than a wrong detail. (GPT Image 2.6)
AI Agents: 97% Approval Is a Trained Reflex
Supervising agents turns the user into the fireground commander, and today’s agent UX trains that commander badly: Anthropic reports that users approve 97% of the permission prompts in Claude Code, its coding agent.
When the company planted a clearly dangerous command in sessions with 1,053 paid testers, only 14% blocked it. The block rate fell from about 17% early in a session to about 5% after 50 or more prompts, whereas an AI classifier blocked 89%. The exact figures deserve caution, because this vendor research took place in a test environment, but the direction matches RPD’s prediction.
Each harmless approval deposits recognition capital in the wrong account: prompt appears, click Allow. Practice made users worse because the cue carried no information. Pavlov’s dogs at least got fed.

When users approve 97% of an agent’s permission prompts, approval becomes a reflex: the bell rings and the paw clicks. Pavlov’s dogs at least got fed. (GPT Image 2.6)
Agents should ask rarely, and only after rehearsing the action. Nobody can imagine at a glance what a chain of commands will do, but the agent can run the rehearsal and show the consequences: the diff, the recipients, the files that will vanish. Agent status should be glanceable, too. When Ruei-Che Chang and co-authors moved it out of the chat log and into ambient cues, the agent’s surviving errors fell by 48% in a 30-user study. (That’s a step toward the audit interface I called for in January.)

Agent status should be glanceable. When researchers moved it out of the chat log and into ambient cues, the agent’s surviving errors fell by 48%. (GPT Image 2.6)
Agents are recognition-primed deciders as well. A model is a veteran of the Internet and a rookie at your company, and it will run the routine playbook on your one non-routine server. Thus, the Kahneman–Klein test applies: valid cues plus feedback, in your environment.

A model is a veteran of the Internet and a rookie at your company. It will run the routine playbook on your one non-routine case. (GPT Image 2.6)


Nobody can simulate a chain of commands at a glance. Let the agent rehearse it and show the consequences: the diff, the recipients, and the files that will vanish. (GPT Image 2.6)
9 Guidelines for Recognition-Primed Design
1. Keep layouts and cue mappings stable across releases. Batch changes, announce them, and budget for a temporary expert slowdown.

Generative UI that redraws the screen for every request banks no recognition capital. Keep layouts stable, or every visit is a first visit. (GPT Image 2.6)
2. Make state glanceable. Fixed positions plus consistent color and shape let experts recognize a situation without reading every label.
3. Make different things look different. Give each cue one meaning, and never let a destructive confirmation share a template with a routine one.
4. Price mental simulation at zero. Offer previews, dry runs, and single-keystroke undo.
5. Scaffold novices separately. Labels, defaults, and progressive disclosure serve users who lack patterns without slowing those who have them.
6. Test with 3–5 real experts at realistic speed. Wherever a veteran hesitates, the design has failed to prime recognition.
7. Make AI show its frame and keep its format. Announce model changes the way you announce redesigns.

A fresh wall of prose must be read, while a constant format can be recognized. Put the verdict in the same place every time. (GPT Image 2.6)
8. Ask for approval rarely and preview the consequences of each request. The dangerous request must look nothing like the routine one.

In Anthropic’s test, human reviewers caught 14% of the dangerous commands, while an AI classifier blocked 89%. Ask people rarely, and only about what matters. (GPT Image 2.6)
9. Make agents rehearse. Require a dry run or a premortem (Klein’s technique: assume the plan failed, then list why) before any irreversible act.

Klein’s premortem is mental simulation for whole plans. Make agents rehearse failure before any irreversible act. (GPT Image 2.6)
Conclusion: The User Is the Fire Commander Now
Recognition is the fastest interaction technique ever invented, and it ships free with every expert. Don’t waste it. As AI moves people from doing the work to supervising it, they’ll decide the way commanders always have: one glance, one option. Design that glance so it lands on the truth.

As AI moves people from doing the work to supervising it, every user becomes the fire commander. Design the glance so it lands on the truth. (GPT Image 2.6)
Final Thought of the Day




