UX Roundup: Boomerang Browsing | Forecasting | Notifications | AI Attitude Gap | Signal Detection Theory | Supporting Evidence | User Feedback | Disabled Buttons | Embodied Interaction
Summary: Misleading information scent sends users boomeranging back to the search results | Forecasters keep underestimating AI, except in the physical world | Don’t swamp users with notifications | Optimistic Asia adopts AI faster than the worried West | False alarms teach users to ignore real ones | Put the evidence beside AI answers | The feedback that trains AI needs usability testing | Disabled buttons are a teachable moment | Users think with their bodies | Keep a human one tap away from the support bot | Replace blank prompts with concrete starting points
UX Roundup for October 23, 2026 (GPT Image 2.5)
Bad Information Scent Boomerangs Users Back
The link promised café recommendations. The landing page coughs up a $2,400 industrial espresso grinder. Seconds later, the user is back at the search results, sniffing for a better link. This is boomerang browsing, and information foraging theory explains why it happens: users are informavores chasing information scent, and when the landing page’s scent contradicts what the link promised, the optimal forager abandons that patch instantly. A fox that smells rabbit but finds an empty burrow doesn’t linger. Neither do users.
Search engines notice these instant returns, so weak scent likely costs you both the visitor and your ranking. Deliver exactly what the link promised, or users won’t stay to browse. They’ll boomerang.
Boomerang browsing: sniff, click, bounce back because the destination didn’t provide what the link promised (smelled like), repeat. Eventually, the user abandons sites with weak information scent for good. (Muse Image)
Forecasters Keep Underestimating AI, Except in the Physical World
Josh Rosenberg and colleagues at the Forecasting Research Institute (FRI) have checked 4 years of AI forecasts against reality. The verdict: domain experts and superforecasters alike consistently underestimate AI capabilities. (Superforecasters are generalists with proven track records.) AI achieved gold-medal performance at the International Mathematical Olympiad (IMO) in July 2025, 5 years before the median expert expected and 10 years before the median superforecaster did.
The lowballing survived the ChatGPT shock: forecasts made as recently as this summer undershot, too.
Even updated forecasts made after AI’s importance became obvious still erred on the timid side. (GPT Image 2.5)
Are labs just teaching to the test? The money says no. Forecasters expected the top AI company to reach $16–25 billion in annualized revenue by the end of 2026. It likely hit $100 billion by September: 4–6 times the forecasts. Customers pay $100 billion a year for useful work, not for benchmark scores.
Biased Scoring Can’t Explain the Misses
FRI candidly flags a bias: a too-low forecast is exposed the day the milestone falls, but a too-high one survives until its deadline. Some unresolved questions were scored against GPT-5.5 projections, and medians hide the more accurate minority. Still, no scoring quirk explains IMO gold arriving 5 years early.
In Greek myth, Cassandra foresaw the future accurately, but Apollo cursed her so that nobody believed her. Some forecasters suffered a milder version of her fate: they were right, but as a minority, they barely moved the median. (GPT Image 2.5)
My favorite detail is pure UX: FRI suspects its survey form made the latest AI revenue figure easy to overlook, so both groups’ median year-end forecasts fell below the revenue already on record. Even forecasting is a usability problem.
The survey design may have made the newest AI revenue figure easy to overlook. Like all users, forecasters rely on whatever the form puts in front of them. (GPT Image 2.5)
Bits Beat the Forecasts; Atoms Don’t
But the errors ran the other way in the physical world. Robotaxis look set to reach about a third of the ride-hailing share experts predicted: they’re safer than human drivers, but operating permits only trickle out. And in a randomized trial, only 5.2% of STEM undergraduates completed three biology lab tasks with AI help, against forecasts of 16–40%. (Data-center forecasts, where construction happens in plain sight, look about right.)
AI software moves at bitspeed, but AI hardware only moves at atomspeed: you can’t copy-paste a robotaxi; somebody has to build each one. (GPT Image 2.5)
Robotaxis are spreading more slowly than forecast, a costly delay given how many lives they could save by reducing traffic accidents. (GPT Image 2.5)
My reading: in the world of bits, forecasters are linearsighted. They draw the future with a ruler, but AI progress only looks straight on log paper. In the world of atoms, they make the opposite mistake, expecting cars and lab benches to scale like software.
Exponential growth curves like AI’s capability gains lie outside everyday human experience, so many forecasters are naturally linearsighted. (GPT Image 2.5)
Get Bifocals
Use the distance lens for AI capabilities. Put an expiration date on every “AI can’t do that” assumption in your roadmap, and recheck it quarterly. (METR’s data show the length of software tasks AI completes autonomously doubling about every 4 months since 2023.) Design for the model you’ll have at launch. By then, the one you tested last spring will be a museum piece.
If you think AI can’t do something, think again. At least once a quarter. (GPT Image 2.5)
Aim where the models will be when you ship. Progress doesn’t pause while you build. (GPT Image 2.5)
Use the reading lens for humans, hardware, and enterprise operations. Model capability isn’t user capability, as those undergraduates discovered with AI at their elbow. Turning one into the other is our job, and it takes longer than any demo suggests.
User capability lags far behind AI model progress. Bridging the gap is the job of forward-deployed designers. (GPT Image 2.5)
Notify Only When It Matters
Aesop covered this: cry wolf over every dessert coupon, and nobody comes running when the real wolf shows up. Each trivial alert (deal of the day, weekly digest, somebody liked something) trains users to swat notifications away unread, like mosquitoes, so when your important message finally arrives, the same reflex flattens it, too.
Attention is a budget, and every interruption costs far more than the few seconds it takes to read. Gloria Mark and colleagues at UC Irvine found that information workers needed an average of 25 minutes to resume interrupted work (PDF), typically detouring through two other tasks on the way back.
So ration alerts like the scarce resource they consume: if it doesn’t help the user right now, it’s an ad. Send fewer. Matter more.
The spray bottle is labeled “Focus.” The swarm remains unimpressed. (GPT Image 2)
93% vs. 36%: The AI Attitude Gap Will Hand Asia the Future
In Gallup’s 37-country survey, 93% of Chinese respondents who know about AI expect it to help their country, vs. 36% of Americans. Optimists adopt, and adopters win, so Asia may pull ahead. The West’s hope: daily AI users worry less than dabblers.
Attitude decides who pulls ahead: optimistic countries adopt AI faster, and adopters reap the gains. (GPT Image 2.5)
Most of the world is curious about AI. The West is worried. So finds a new Gallup survey of 37 countries, with about 1,000 adults each.
Pessimism in the West, optimism in Asia. This one number may dictate the future. (GPT Image 2.5)
Asked whether AI will mostly help or harm people in their country, 93% of Chinese respondents who know about AI said “help,” as did 91% of Vietnamese and 77% of Singaporeans, but only 36% of Americans. Meanwhile, 74% of AI-aware Americans say AI makes them worried, and rich Western countries hold 7 of the top 8 worry spots.
Most Americans worry about AI, but fewer than 3% pay for an AI subscription, so few have firsthand experience with good AI. Instead, they rely on scare stories from legacy media. (GPT Image 2.5)
Optimists Adopt, and Adopters Win
Why does attitude matter? Because AI pays off only for people who use it often. Singapore has 46% daily AI users and China 36%, vs. 23% in the United States. Singapore’s daily-use rate is double America’s. AI is a forklift for the mind, but only for people who drive it daily.
History rhymes, with the roles reversed. In 1793, China’s Qianlong Emperor told King George III that his empire had no use for British manufactures, and China then largely sat out the Industrial Revolution, paying for it over the next century.
Today, the West plays the emperor. My prediction: if these attitudes persist, Asia will reap the biggest AI gains of the coming decade.
China suffered a century of humiliation after its emperor scorned British innovation. Maybe it’s cosmic justice that a similar fate now looms over the UK for worrying too much about AI. (GPT Image 2.5)
Many politicians in the US and Europe are about to repeat the Qianlong Emperor’s 1793 mistake. (GPT Image 2.5)
Daily Use = Less Worry
Now for my hope. Within the West, heavy users worry least. In the United States, 68% of daily users feel worried, vs. 80% of occasional users (who use AI less than weekly), and 45% of daily users trust AI’s accuracy, vs. 8% of nonusers.
Often, dabblers worry the most: in 4 of the 7 Western countries, occasional users worry more than people who have never touched AI. Dabbledread sets in after a few clumsy sessions: you see the hallucinations but quit before AI turns your mediocre draft into a good one. It’s a toe in cold water: all shock, no swim. And America is a nation of dabblers: 24% of adults use AI less than weekly, the highest share of all 37 countries and more than the 23% who use it daily.
Monthly AI dabbling works like a monthly gym visit: all soreness, no strength. (GPT Image 2.5)
Gallup rightly notes that a survey can’t tell whether use builds trust or trust drives use. (Probably both.) But in randomized experiments by Yaoxi Shi and co-authors, people assigned to get emotional support from AI grew more likely to choose AI next time, even if they had wanted a human. (My very first newsletter reported the user-vs.-nonuser gap in 2023.)
Does use build trust, or does trust encourage use? Probably both, but people who never use frontier AI are bound to misjudge it. (GPT Image 2.5)
The benefits are real: AI uplevels human potential and makes people more creative, especially the least creative. Nobody discovers this in a monthly session.
Unfortunately, many AI labs keep talking about AI’s potential risks, even though those risks are either purely hypothetical (AI will kill us all) or much smaller than equivalent risks from non-AI technology (cyberattacks, which AI can help defeat). Sensationalist legacy-media coverage of these hypothetical risks creates fear among people who lack personal experience with AI. (GPT Image 2.5)
The Cure for AI Worry Is AI Use
The West has plenty of AI technology and too few AI users. Asia’s lead is in attitude, and attitudes change with experience. Managers: don’t settle for occasional use; give every employee a daily task where AI clearly helps.
Designers: Zhu and co-authors found that prompting habits crystallize within 3–5 sessions, so make the first sessions pay off fast. Familiarity breeds contentment, but only for users who stick around long enough to get familiar.
Daily AI use for big and small projects, at work and at home: that’s the way to beat the fear. (GPT Image 2.5)
Bottom line: AI is the same, but if you think optimistically about what it can do for humanity, the future will be better. Negative thinking is self-reinforcing. (GPT Image 2.5)
Signal Detection Theory: False Alarms Teach Users to Ignore Real Ones
Every alerting system produces exactly 4 outcomes: hits, misses, false alarms, and correct rejections. Any threshold merely trades one kind of error for the other. Most digital products buy a few avoided misses with mountains of false alarms, and users respond by tuning out the whole channel. Treat false alarms as withdrawals from a finite trust account.
One true alarm bell among a thousand maybes: the observer’s 4 possible verdicts (hit, miss, false alarm, correct rejection) frame every notification, warning, and spam filter you’ll ever ship. (GPT Image 2)
Definition: Signal detection theory (SDT) models decisions in which a genuine signal must be distinguished from background noise. Its central move is separating sensitivity (how well signal and noise can be told apart, written d′) from the response criterion (how much evidence the decision-maker demands before declaring “signal”).
Radar crews, spam filters, and smoke detectors all live in this 2×2 grid, and 2 of its 4 cells are errors. (GPT Image 2)
Born on Radar Screens
Is that blip a bomber or a flock of birds? World War II radar crews faced the question nightly, and one morning it mattered more than most. On December 7, 1941, the radar station at Opana Point, Hawaii, picked up a large formation inbound; the crew reported it; the duty officer waved it off as a flight of B-17s expected from California.
The equipment scored a hit. The decision chain converted it into the most consequential miss of the century.
After the war, Wesley Peterson and colleagues at the University of Michigan formalized the mathematics of signal detectability in 1954, and that same year Wilson Tanner and John Swets imported the framework into psychology with “A Decision-Making Theory of Visual Detection” in Psychological Review. David Green and John Swets’s 1966 book Signal Detection Theory and Psychophysics completed the takeover, and the theory now runs through diagnostics, spam filtering, fraud detection, and content moderation. The name is admirably literal: detecting signals. In noise.
Two Dials, One Honest Trade-Off
An alerting system has only two dials. The first is sensitivity: improve the underlying evidence (better data, better sensors, more context), and both error types shrink together. The second is the criterion: slide it liberal, and you trade misses for false alarms; slide it strict, and you trade back. There’s no third dial. Teams promising “zero missed fraud and zero false declines” are selling perpetual-motion machines.
Choosing a criterion means comparing costs, and the comparison has a term most teams omit: every false alarm also debits future attention.
Medicine measured the wreckage. The Joint Commission’s 2013 Sentinel Event Alert on alarm safety (PDF) reported hospital units generating hundreds of alarm signals per patient per day, with an estimated 85–99% requiring no clinical intervention. Clinicians became desensitized, alarms got silenced or detuned, and patients died. If trained nurses learn to ignore life-critical alarms, your app’s little red badge doesn’t stand a chance.
(Usability evaluation is a detection task, too: a single inspector finds only about 1/3 of the real problems in an interface and flags some non-problems along the way, which is why I’ve recommended pooling 3–5 evaluators since the early 1990s.)
Spend Your Cry-Wolf Budget Like Cash
Every product holds a cry-wolf budget: the finite number of false alarms users will forgive before reclassifying the entire channel as noise. Aesop published the core finding roughly 2,600 years before Tanner and Swets; his shepherd ran the experiment and lost the sheep.
Here’s the standard failure mode. Each miss-averse stakeholder (security, legal, growth, “engagement”) sets a liberal criterion for his or her own alerts, and the union of all those defensible decisions lands on the user’s lock screen as a wall of noise.
So users reach for the only criterion they control: the global mute. But a global mute is a designed miss for everything, including the one alert that will someday matter. And alarmspam, marketing pings dressed up as warnings (“Your cart is in danger!”), burns the same budget at a spectacular rate.
Escape routes exist. Raise d′ before touching the threshold: add context such as location, device history, and severity scoring, so that alerts fire on better evidence rather than louder hunches. Respect base rates: when the true event is rare, even an accurate detector produces mostly false positives, so rare-event alerts must be conservative, or users will ignore them by week 2.
7 Design Guidelines for Alerts and Warnings
Keep the wolf call believable:
Decide, in writing, which error costs more for each alert type, and set its threshold from that decision instead of from stakeholder anxiety.
Improve the signal before amplifying the alarm. Better evidence shrinks both error types; louder merely reallocates the annoyance.
Enforce severity tiers. Interruptive modalities (modals, sound, vibration, red) belong to the act-now tier alone; everything else waits its turn.
Make every alert actionable. State what happened and what the user should do next; an alert without a next step is noise with good posture.
Digest the low tier. Batch minor events into a periodic summary that users read on their own schedule.
Track dismissal rates as a KPI. An alert ignored by 95% of recipients has measured its own worth, so demote or retire it.
Offer per-category controls, because the only alternative you leave users is the global mute: total blindness by design.
The boy who cried wolf had a signal problem: by the time the real wolf came, his shouts carried no information. Most of the time, silence is a feature. Guard yours.
Alice and Zimo explain signal detection theory in AAA Cinematic 3D style. (GPT Image 2)
Put the Evidence Beside the Answer
A polished AI answer can look finished before anybody has inspected its factual foundation. Confident language, clean typography, and an orderly explanation make a persuasive pedestal. Open the little door at its base, though, and the evidence compartment may be empty.
An answer can look rock-solid while its foundation is hollow. (GPT Image 2)
Telling users to verify everything dumps a fresh research project on people who came to the AI to save work. The interface should lighten that burden wherever factual support matters, especially for claims that will drive a consequential decision.
In my discussion of making checking cheaper than trusting, I recommended opening the exact supporting passage beside the AI’s claim. Side-by-side evidence lets users compare what the source says with what the model concluded, without losing their place.
Attach references to the claims they support. Distinguish retrieved facts from inference, and make unsupported assertions visible. A link to a long document’s first page is a treasure map without the X: users must still dig for the passage that bears the argument’s weight.
To test your design, pick one important factual claim and ask users to track down its support. Observe whether they can find the evidence and judge whether it actually justifies the wording. Verification is an interaction to design. The pedestal should open easily, and there should be something inside.
The Thumbs That Train AI Deserve Usability Testing
Benchmarks measure what AI labs care about. User feedback measures what users care about. The thumbs under AI answers matter because the votes become training data. But votes can also mislead: thumbs data helped make OpenAI’s 2025 GPT-4o update sycophantic. A thumbs-up can’t tell accuracy from flattery.
A case full of benchmark trophies can’t tell you whether the AI met user needs. User feedback can. In fact, I suspect that, outside of coding, AI labs mostly chase the wrong benchmarks. (GPT Image 2.5)
AI products almost always ask why after a thumbs-down but rarely after a thumbs-up. Positive presets such as “accurate” or “saved me time” tell the model what to keep doing. (GPT Image 2.5)
Hans Christian Andersen described the problem in 1837: reward flattery, and nobody tells the emperor he’s naked. OpenAI named thumbs data as one cause of its sycophantic 2025 GPT-4o update. (GPT Image 2.5)
So I’m glad Alisa Frik and co-authors from the International Computer Science Institute and eBay usability-tested these humble widgets. They reviewed feedback features in 40 AI products, had 24 eBay experts use my classic 10 usability heuristics to run a heuristic evaluation of 12 of them (a method I’m partial to, having co-invented it), and tested a prototype with 13 eBay users.
Of 40 popular products with generative AI features, 8 offered no feedback mechanism for the AI at all. (GPT Image 2.5)
Feedback buttons hidden behind hover, long-press, or a menu might as well be wrapped by Christo and Jeanne-Claude. Show the thumbs right below the output, no hovering required. (GPT Image 2.5)
Thumbs up/down beat a 5-star scale, which one user dismissed as “too many shades of gray.” But users disagreed about what the thumbs rated: the text, the recommended products, or the whole feature. Such thumbiguity muddies the training data, so label the scope (“Share feedback on image”). Users liked a compact menu of about 8 quick preset reasons in plain words. Presets also capture why: “accurate” and “liked the tone” are different compliments.
Thumbiguity is 2,000 years old. When one answer holds text, images, and product picks, label each set of thumbs with what it rates, such as “Share feedback on image.” (GPT Image 2.5)
Users preferred thumbs to a 5-star scale. Optometrists learned long ago that “better or worse?” is easier to answer than “how good, on a scale of 1 to 5?” (GPT Image 2.5)
Thank users for their feedback, but keep it short and sincere. A thank-you so effusive that it sounds fake undoes the goodwill. (GPT Image 2.5)
Caveat: this study is hardly definitive, since 13 users suffice to find usability problems but not to crown winners among a handful of design variants. Still, kudos to eBay for applying classic usability methods to one small piece of AI UX. We need much more of this. Meanwhile, steal the paper’s feedback template.
The paper’s appendix holds a complete feedback template, free for the taking. Use it until you have time to run your own user research and design something better. (GPT Image 2.5)
Disabled Buttons Must Explain Themselves
A grayed-out button is an interface shrugging at its customer. The validation code knows precisely which field is missing or malformed; the pixels refuse to pass the information along. So the user resorts to archaeology: rereading every field, retyping the phone number in three formats, hunting for the unchecked box. Some give up mid-purchase, which converts a programming shortcut into lost revenue.
Two better patterns exist. Option one: leave the button enabled, and when it’s clicked prematurely, respond with specific, polite error messages pinned to the offending fields. Option two: when a GUI control is disabled, print the reason next to it (“Enter a phone number to continue”), updated live. Either way, the system must never know more about the problem than it tells the user. A disabled control is a teachable moment, so spend it telling the user exactly what to do.
Silence can pass for minimalism, but a control that withholds its own requirements is a locked door with the keyhole painted over.
The interface knows exactly why you can’t continue; it’s just not telling. (GPT Image 2)
My music video: the disabled button blues (YouTube, 4 min.).
Embodied Interaction: Users Think With Their Bodies
Users think with their hands, eyes, and muscle memory, and interfaces succeed when they recruit those ancient skills. Embodied interaction explains why direct manipulation works and why gesture-only interfaces fail. Keep feedback under 0.1 seconds, give every gesture a visible twin, and always budget the body tax.
Each object answers where he touches it. That closed loop between body and system is the point: the whole posture becomes the input device, and the output is something you can feel your way along. (GPT Image 2)
Meaning Lives in the Doing
Definition: Embodied interaction is the view that people create and grasp meaning through engaged physical and social action in the world, and that interfaces therefore work best when they recruit skills users already own: reaching, grasping, spatial memory, turn-taking.
Want proof that the body computes? David Kirsh and Paul Maglio of the University of California, San Diego, studied Tetris players and found that experts rotate falling pieces on screen far more than any efficient plan requires, because physical rotation is faster and more reliable than mental rotation (“On Distinguishing Epistemic From Pragmatic Action,” Cognitive Science, 1994, PDF).
They called these moves epistemic actions: physical actions that make thinking easier, faster, or more reliable. Your users do the same when they drag files into piles or zoom a map in and out to get their bearings. The hand is doing cognition the brain would do worse.
Philosophy That Escaped the Seminar Room
Paul Dourish of UC Irvine named the field in Where the Action Is: The Foundations of Embodied Interaction (MIT Press, 2001), braiding two 1990s research strands: tangible computing, launched by Hiroshi Ishii and Brygg Ullmer’s “Tangible Bits” (CHI 1997), and social computing, rooted in Lucy Suchman’s Plans and Situated Actions (1987).
The name is a rebuke. Classical HCI inherited from 1970s cognitive science a model of the user as a disembodied information processor, a brain in a vat with a mouse attached. “Embodied” insists that interaction is something bodies do, in places, among people.
The philosophical scaffolding is older. Martin Heidegger’s Being and Time (1927) observed that a hammer in skilled use is ready-to-hand, effectively invisible, until it breaks and becomes present-at-hand, an object you must stare at. (Every crashed app performs a small Heidegger demonstration.) Maurice Merleau-Ponty’s Phenomenology of Perception (1945) added that the body is our medium for having a world at all.
Ben Shneiderman needed neither the Germans nor the French: his 1983 article “Direct Manipulation: A Step Beyond Programming Languages” named the technique, and GUIs conquered the desktop because dragging a document to the trash recruits hand-eye skills every user has honed since infancy.
The Body Is a Free Peripheral
Treat the user’s body as pre-installed hardware and design to its specs:
Muscle memory and spatial memory. Stable layouts let hands find controls without visual search, the way a concert pianist rarely glances at the keys. Keep the Send button where it stood in the last release, and the user’s fingers file the paperwork while his or her mind stays on the message.
The 0.1-second loop. In my 1993 book Usability Engineering, I set 0.1 seconds as the response limit for feeling that you, and nobody else, caused the on-screen effect. Below that limit, screen objects feel physical; above it, the loop breaks and “direct” manipulation degrades into remote control.
Thinking by doing. Cheap, reversible manipulation (drag to rearrange, pinch to zoom, scrub a timeline) lets users offload cognition onto the screen, like the Tetris experts. The iPhone’s 2007 masterstroke was making content itself the touched object, firing the cursor as middleman.
Recede on purpose. The highest compliment an interface can earn is to go unnoticed. Ready-to-hand is the goal state, so log as a defect every moment that forces users to notice the tool.
When Embodiment Becomes a Gimmick
The failure pattern arrived with the touchscreen boom. Vendors declared gestures “natural” and shipped interfaces with no visible controls, no discoverable commands, and no consistency across apps. Don Norman and I called it out in “Gestural Interfaces: A Step Backward in Usability” (interactions, 2010): the rush to gesture threw away visibility, feedback, and undo, the hard-won principles that make skilled bodily use possible in the first place. Shneiderman had promised a step beyond; too many gesture UIs delivered a step backward. A gesture with no signifier is no more natural than a secret handshake.
Then there’s the body tax: the physical cost an interface levies on muscles and joints. Vertical touchscreens invite gorilla arm, the fatigue that engineers have joked about since the light-pen era of the 1960s. Microsoft’s Kinect (launched 2010, discontinued 2017) asked living-room players to flail for hours, and VR headsets add simulator sickness to the invoice.
Worse, naive embodiment assumes one ideal body: two steady hands, full range of motion, standing still. Real users hold the phone one-handed on a moving train, type with a tremor, or reach with a broken wrist. Design that recruits the body must also forgive the body.
The remedies are unglamorous, and they work: signifiers and hints for every gesture, a visible button path in parallel, platform-standard gestures only, feedback within 0.1 seconds, and testing with diverse bodies in awkward postures.
7 Design Guidelines for Embodied Interaction
Give every gesture a visible twin. Ship a button or menu path for each gesture command. Gestures make fine shortcuts for experts and terrible front doors for everybody else.
Close the loop in 0.1 seconds. Touch, drag, and pen input must respond within 0.1 s, or the sense of physical causation dies.
Freeze the spatial layout. Muscle memory is an asset users build for you free of charge; a redesign that shuffles controls burns it in one release.
Use standard gestures, and always signify them. Stick to the platform’s gesture set, add visual hints (handles, partial reveals, first-run coaching), and back every gesture with undo.
Budget the body tax. Ban sustained mid-air or above-shoulder input, and run 20-minute test sessions that log reported fatigue alongside errors.
Support thinking by doing. Make objects draggable, zoomable, and rearrangeable with instant reversal, so users can act to think.
Test with real bodies in real postures. Include one-handed use, left-handed use, gloves, tremor, and transit conditions; an interface that demands an ideal body excludes paying customers.
Alice and Zimo explain embodied interaction in AAA Cinematic 3D style. (GPT Image 2)
Keep a Human One Tap Away
Contact centers grade their chatbots on containment rate, the share of customers who never reach a human. Read that metric again: the customer is the thing being contained. Optimize for it, and you get a bot whose real job is deflection. It makes customers restate the problem, offers the help article they already read, and keeps the door to a human bolted shut. Gartner surveyed 5,728 customers in late 2023: 64% would rather companies didn’t use AI for customer service, and 53% would consider switching to a competitor over it.
Their top fear, named by 60%, is that AI makes it harder to reach a human.
I’m an AI optimist, so note what that survey indicts: bots hired as bouncers. A good bot resolves the easy cases and hands off the rest within 2 exchanges, transcript attached, so nobody has to retell the story. A support bot that can’t escalate is a moat around the support team. Put the human one tap away, and the bot can earn its bow tie.
Containment rate: 100%. Customers contained: all of them. (GPT Image 2)
A Blank Prompt Is an Empty Menu
“Ask me anything” is a peculiar sales pitch. It promises limitless assistance while requiring users to invent the service they came to receive. Picture a waiter who announces, “The kitchen can make anything,” and then stands there, pen poised, waiting for you to invent dinner.
A real menu shows diners what the kitchen can prepare. An AI interface should likewise reveal useful possibilities in words that match tasks people recognize from their own work. Otherwise, users must guess both what the system can do and how to ask for it. My article on prompt augmentation explores ways to replace this guesswork with recognizable choices.
Concrete, editable starting points help users discover useful AI capabilities without having to invent their first request. (GPT Image 2)
Consider a new employee opening an internal AI assistant. “Draft my reply to this customer,” “Explain this expense policy,” and “Find the latest project decision” offer concrete beginnings. Each example should open an editable request, indicate the information needed, and let the employee inspect it before submitting.
Good starting points teach through use. Completing a worthwhile task gives users an example they can adapt for the next one. A wall of prompt-writing advice merely assigns homework before the work.
Choose three common tasks and test whether new users can recognize one relevant to them. Then watch whether they reach a useful result. Clicking a suggestion proves little if it leads into a dead end.
Keep the open prompt available for unusual requests and for users who already know what they want. Freedom becomes more valuable once people understand the possibilities.
An empty menu gives diners unlimited choice and remarkably little chance of ordering lunch.
Final Thought of the Day
