UX Roundup: Human “Taste” Less Unique | World Usability Day | New Work With AI | Agent Waits | Pale Text | AI Teammates | Evidence | Proximity Compatibility Principle | Back Button
Summary: Taste stops being a unique human advantage | This Thursday is World Usability Day | AI speeds up old work and unlocks new work | What to do while waiting for AI agents | If users can’t read it, it doesn’t exist | AI teammates cut the labor cost per ad by a third | Evidence wins design arguments | Display proximity should match processing proximity | Don’t let Back wipe the user’s filters and scroll position
UX Roundup for November 9, 2026 (GPT Image 2.6)
The Taste Barrier Falls: One AI Prompt Out-Designs Professional Creatives
In a market experiment with 1,025 people, games designed by AI from a single prompt beat amateurs’ games by 27% and topped professionals’ games, too. Revealing human authorship didn’t increase consumers’ willingness to pay. For these small games, the taste barrier has fallen; the human touch commanded no premium.
AI goes up against experienced game designers and wins. (GPT Image 2.5)
AI first beat humans at playing games: chess in 1997, then Go in 2016. Now it beats us at designing small games as well.
AI’s history shows it beating humans at progressively harder challenges. “Taste” is the next target. (GPT Image 2.5)
John Conlon (Ohio State University) and Peter Schwardmann (Carnegie Mellon University) built a miniature market to test two hopes for preserving human creative work: superior taste and a premium for the human touch. In this experiment, neither hope survived.
One Prompt Beat Three Rounds of Human Direction
In the study, 296 amateurs and 180 professional creatives (game developers, graphic designers, and content creators) improved simple black-and-white games through three rounds of instructions to an AI coding agent, playtesting between rounds.
The AI worked with a handicap. It got one standard prompt per game, asking it to make the game “much more creative, engaging, and fun to play.” Its single reply went straight to the same coding agent, without any playtesting.
Consumers bid for the chance to play each game, and the game’s producer earned the amount bid.
Gamers were willing to pay more for an AI-designed game than for the work of human designers. (GPT Image 2.5)
AI-designed games fetched $0.47, compared with $0.37 for amateurs’ games, a 27% advantage. Professionals did better than amateurs but still lost, and consumers rated the AI’s games more fun, more polished, and, ouch, more creative.
Critics say AI games and other media, such as movies, lack soul. When consumers prefer AI content, businesses will supply it, soul or no soul. (GPT Image 2.5)
The AI was an idea fountain, with a typical brief of 235 words, compared with 60–89 words for the humans across all three rounds combined. One AI turn produced more direction than three human turns.
Designers love to complain about vague client briefs. In this study, the humans wrote the vague briefs, and the AI wrote the detailed one. To be fair, the humans wrote more than “make it pop,” but only about a third as much as the AI. (GPT Image 2.5)
Human Touch = Zero Premium
My 18th prediction for 2026 was “Human Touch as Luxury: No.” I allowed for a temporary premium during the transition, but even that premium was a no-show in this experiment.
Telling consumers that a human designed the game had no measurable effect on its value. (GPT Image 2.5)
Revealing who designed each game left the AI’s advantage intact: 6.8 cents when authorship was hidden versus 6.0 cents when it was revealed (p=0.79). The advantage also held when consumers were told that their money would go to the human creator.
Without labels, consumers couldn’t even spot the human games. They rated those games 58% likely to be human-made, compared with 59% for the AI’s games.
Consumers who trusted their gut to detect the human touch did no better than a coin toss. (GPT Image 2.5)
In this market, customers paid for quality; the creator’s identity added nothing.
The market prizes the blade over the blacksmith. (GPT Image 2.5)
Taste: From Prediction to Proof in 8 Months
In February, I wrote that the bitter lesson was getting more bitter for believers in “meatware supremacy.” AI’s judgment improves with compute and, I predicted, would likely surpass human “taste” (or rather, design judgment) within a few years.
Each new study brews the bitter lesson a little stronger for believers in meatware supremacy. (GPT Image 2.5)
For small games, reality caught up with that prediction in 8 months. No elite band of human designers held out against the AI, whose games led at every price decile except the bottom one.
Design judgment was supposed to be a human-only ability. For small designs, AI has it too. For larger projects, humans will probably keep the lead for another year or two. (GPT Image 2.5)
Hundreds of human designers competed, and not even the top tier beat the AI’s best games. (GPT Image 2.5)
The designers never saw their defeat coming. Amateurs expected to beat the AI by 9 cents and professionals by 17 cents, so the pros had the worst tastefidence: overconfidence in their own taste.
Professionals expected to beat the AI by 17 cents and changed their minds the moment they read its idea. That’s tastefidence. (GPT Image 2.5)
Both groups reversed course as soon as they read the AI’s idea. They recognized a better idea on sight but couldn’t produce one themselves. Watching the AI improved their later games and deflated their pride.
Human designers judged the AI’s ideas well. They just needed to see them first. (GPT Image 2.5)
Learning from the AI made for better games and humbler designers. (GPT Image 2.5)
Granted, these were one-minute games with bids measured in cents, and the professionals came from an online panel. The setup still favored the humans in one respect: the solo AI designer, Claude Sonnet 4.6, was a midrange model. As AI’s task horizon grows, I expect larger designs to fall next.
Taste was supposed to be the moat that kept AI out of creative work. For small designs, the moat has already dried up. (GPT Image 2.5)
Conclusion: Move Upstream from Taste to Agency
Don’t bet your brand on a “human-made” label. Before designing anything, generate an AI-only version. Ship yours only if it beats the AI’s version with real users. Your own preference doesn’t prove that your design is better. That’s tastefidence talking.
Then invest your time where the study’s producers had no say: deciding what to build, whom to build it for, and why it deserves to be built.
When AI handles the design, humans can focus on choosing what deserves to exist. (GPT Image 2.5)
Chess, Go, and now small games. Your move.
World Usability Day: November 12
This Thursday, November 12, 2026, is World Usability Day. Check the website for events in your country. Or celebrate the classic way, by testing with 5 users and fixing what trips them up.
Happy World Usability Day! Let’s unlock a better world through better usability. (GPT Image 2.6)
Faster Work ≠ Fewer Jobs: AI Also Unlocks New Work
Almost half (45%) of what developers say AI does for them is old work done faster. Rudrajit Choudhuri and co-authors from Oregon State University and Microsoft asked 331 Microsoft developers what AI lets them do that they couldn’t do before.
Another 37% of answers described working through the wishlog: work that had once been beyond the developers’ skills or hadn’t been worth the effort. A further 12% described brand-new work, including directing AI agents, writing the standards that guide them, and building for the agents themselves.
Self-reported speedups need a stopwatch. In an earlier METR study, developers believed AI sped them up while it actually slowed them down by 19%. (GPT Image 2.5)
The hours AI saves don’t vanish. Developers spend them on work they always wanted to do but never could. (GPT Image 2.5)
In 12% of answers, developers described brand-new kinds of work, such as directing AI agents. Somebody has to hold the leashes. (GPT Image 2.5)
Building for AI agents is new work, too. The agents now get the onboarding documentation that human hires always wished for. (GPT Image 2.5)
Doomsayers fear the speedup because they assume that the amount of work is fixed: if the same work takes fewer hours, companies must need fewer people. That’s the fixed-work fallacy I debunked in 2023.
When productivity reduces the labor needed for old tasks, people can tackle work they had previously set aside and invent new tasks as well. Within this single workforce, answers about new work outnumbered speedup answers, 49% to 45%. Faster work doesn’t automatically mean fewer jobs.
After years of warnings about AI-driven mass unemployment, the clerks still have time to knit. (GPT Image 2.5)
Farming didn’t leave 98% of hunters jobless, and factories didn’t strand 98% of farmers. AI won’t strand today’s knowledge workers, either. (GPT Image 2.5)
The authors liken grading AI on speed alone to grading electricity on how much it sped up steam-era factories. Electricity changed far more than the speed of existing work, and it didn’t cause mass unemployment, either.
The word “computer” used to be a job title for the workers (mainly women) who computed ballistic tables for the Army. After electronic computers arrived, many human computers became programmers. (GPT Image 2.5)
In 1954, IBM’s first FORTRAN report predicted that the language would “virtually eliminate coding and debugging.” Today, the world employs more than 20 million professional programmers (plus about as many hobbyist programmers). (GPT Image 2.5)
To claim the new work, grow your judgment as fast as AI grows its skills. Otherwise, you’ll become what the authors call a hollow orchestrator: someone who approves output he or she can no longer evaluate.
While AI Agents Work, Review the Decisions You Never Made
Minute-long agent runs are too long to watch and too short to leave. In a study, a read-only side chat that follows the agent’s run cut users’ time on other tabs from 35% to 13% of the wait and surfaced decisions they had left unmade. But users paid for it with higher mental demand.
In Slow AI, I tackled agents that run for hours: users leave, so design for their return. Quick runs earned a shrug from me, since you could always retry a failed one. Those retries may cost few tokens, but they still drain the user’s attention.
Batch processing was miserable but honest: you knew you could go home. A micro-wait keeps you tethered without keeping you busy. (GPT Image 2.5)
Kevin Pu and co-authors from the University of Toronto and Google DeepMind call these minute-long waits micro-waits: “too long to watch, too short to leave.”
Their 12 interviewees, all daily agent users, faced a lose-lose choice. They could stare at an inscrutable reasoning trace, or leave and risk drifting into other tasks or social media, losing the context needed to re-engage.
One interviewee said a run might finish “within five seconds, or like an hour.” The researchers tried using AI to predict the wait, but its forecasts weren’t accurate enough to use. (GPT Image 2.5)
Leaving a micro-wait creates a second watch duty, because the agent may finish, stall, or ask for input at any moment. (GPT Image 2.5)
Switch away long enough, and you return with no memory of your own request. That’s why hour-long runs need resumption summaries. (GPT Image 2.5)
Reasoning traces show everything and explain little. Users need a clear summary of the key decisions. (GPT Image 2.5)
Watching the trace catches the occasional bug, but at a dismal hourly rate. (GPT Image 2.5)
Sidecar = A Seat Without Handlebars
The team’s prototype, Sidecar, is a side chat in the agent’s browser tab. Like its motorcycle namesake, it rides along without steering. It reads the whole run but can’t touch the work. It lists the agent’s unrequested decisions, suggests directions to explore, and lets users draft their next prompt while the agent keeps working.
Sidecar aimed to help users digest the agent’s reasoning trace. Most humans prefer a summary. (GPT Image 2.5)
In a follow-up study, 16 participants prototyped web apps or generated videos, with and without Sidecar. Median waits ran 72–76 seconds, more than 7 times the 10-second attention limit I published in 1993.
With Sidecar, time spent on other tabs fell from 35% to 13% of the wait. Median ratings of progress during waits rose from 2 to 5 on a 7-point scale. But median mental demand also doubled, from 2.5 to 5.
Midsight Beats Hindsight
The big win is midsight: 6 of 16 participants reported catching unwanted agent decisions during the run rather than discovering them in the finished output.
Sidecar’s cards exposed holes in users’ own prompts, including what one participant called “places where I had to make decisions that I didn’t think about.” The decision cards are my conceptual breadcrumbs from Slow AI, shrunk to fit a minute.
Every prompt leaves decisions open for the agent to fill in. Decision cards surface those choices during the run, so users can prepare a follow-up if the agent has taken a wrong turn. (GPT Image 2.5)
Decision cards work like a warning light for underspecified prompts, flashing mid-run to give you time to prepare your next instruction. (GPT Image 2.5)
(Caveat: Google researchers tested Google’s own agent platforms with big-tech staff for about 15 minutes per task. Still, the tab-switching data comes straight from logs rather than from participants’ recollections.)
What to Do During Agent Waits
Without Sidecar, do its job by hand. While the agent works, ask yourself which decisions your prompt left open, and draft the follow-up you may need.
Use the wait to draft your next prompts, so they’re cleared for takeoff the moment the agent finishes. (GPT Image 2.5)
Designers should build a side lane:
Keep wait-time help in the agent’s tab, with the run’s full context.
Put decisions front and center; keep traces out of the way.
Make the side lane read-only, so half-baked ideas can’t trigger action.
Main agents treat brainstorming as a work order. Users need somewhere to think without triggering construction. (GPT Image 2.5)
Ideas arrive mid-run, but sending them would interrupt the agent. Pelicans have a pouch for their catch; users need one for half-formed prompts. (GPT Image 2.5)
Stay one step ahead at a calm pace. In the study, suggestions that went further or arrived faster raised cognitive load, adding to the very burden the tool was supposed to ease.
Wait-time help must be ambient and user-initiated. Pop-ups provoke the reaction one interviewee predicted: “don’t tell me what to do.” (GPT Image 2.5)
For hour-long runs, design for the user’s return. For minute-long runs, help the user stay engaged without turning the wait into another job.
Peripheral awareness beats staring. Keep the run at the edge of attention, close enough to notice when it finishes. (GPT Image 2.5)
If Users Can’t Read It, It Doesn’t Exist
Pale gray text on a paler gray background is fashionable, and this “graywash” has a blind spot. Designers evaluate their screens on color-calibrated monitors, at full brightness, often while they’re in their thirties or younger.
Users may read the same text on a 3-year-old phone, outdoors, at half brightness, often with the older eyes that come standard with disposable income. Contrast sensitivity declines steadily with age, so the gray-on-gray look punishes some of your richest customers hardest.
The floor is well defined: WCAG Level AA requires a contrast ratio of at least 4.5:1 for body text and 3:1 for large text. Treat those ratios as a minimum, and aim higher when you can.
The test costs nothing. Open your product on a cheap laptop, near a window, at 50% brightness. If the content fades from view, it fails at its job: unread text has a business value of zero. Let decoration whisper, but make your content speak up.
Text that blends in too well might as well not have been written. (GPT Image 2)
AI Teammates Cut Labor Cost per Ad by a Third Without Losing Clicks
Harang Ju (Johns Hopkins) and Sinan Aral (MIT) randomly assigned 2,234 people to make ads for a think tank’s annual report with either a human or an AI teammate. Most AI creativity studies end at the judges’ table.
Ju and Aral went further and ran 2,000 of the 11,024 ads on X, earning 4.9 million impressions and 7,546 clicks from real users. Their field test measured how the creations performed with an actual audience. Bravo.
What judges say and how designs perform in the real world often differ. (GPT Image 2.5)
Real Clicks Validate the Ratings
AI improved the copy but worsened the pictures, and neither team type’s ads drew significantly more clicks in the field. But the field test validated the ratings: each point of rated text quality (on a 7-point scale) was associated with roughly a 4% lift in click-through rates. That’s the top rung of my three levels of success for AI writing: measuring what recipients do.
The advantage was cost. With an AI teammate, each person produced 50% more ads than a member of a human pair, cutting the labor cost per ad by a third. Cheaper ads, no measurable loss of clicks.
Lower cost alone gives AI a business advantage. (GPT Image 2.5)
With an AI teammate, people produced 50% more ads than with a human teammate. (GPT Image 2.5)
Delegation Breeds Sameness
The AI-assisted ads were also more alike, echoing Anil Doshi and Oliver Hauser’s finding that AI boosts individual creativity while shrinking collective variety. The more copy people delegated to AI, the more similar their resulting ads became.
Samesourcing your creative work to the AI everybody uses gets you everybody’s ideas. The next research frontier is AI that widens the spread of ideas instead of funneling creators toward the same answer. (For now, prompt your AI to generate a wide variety of options, as I did for the cartoons in this news item.)
The more people delegated to the AI, the more their ads converged on the same shade of beige. Variety needs a design goal of its own. (GPT Image 2.5)
If you don’t ask AI for variety, you may get only its best guess at a solution. (GPT Image 2.5)
Amateurs With Antique AI
Two caveats temper these results. First, although the ad makers were a representative sample of Americans paid $9 for 40 minutes, real ads come from agency creatives, small-business owners, or marketing staff. Doshi and Hauser found that AI helped the least creative writers most, so I suspect that professionals gain less from AI assistance.
The study’s participants were ordinary Americans with no regular ad-creation experience. Its results are more likely to generalize to very small businesses than to professional advertising agencies. (GPT Image 2.5)
Second, the AI is obsolete. The study ran in October 2024 with GPT-4o as the teammate and DALL-E 3 (from 2023) as its image generator. Both were pioneering models in their day, but they look feeble by today’s standards.
In AI years, that’s the Bronze Age. The image-quality finding therefore says little about what current tools can do.
DALL-E 3 was announced in September 2023, which makes it an antique as a study’s image model. Its poor image ratings come as no surprise: those were the days of extra fingers. (GPT Image 2.5)
Academic publishing lags so far behind AI that many papers are archaeology by the time they appear. (GPT Image 2.5)
Boss the Bot, Don’t Befriend It
With an AI teammate, people sent 62% more messages and made 62% fewer direct edits, shifting from editing to directing. Task talk predicted more and better ads; social chitchat predicted fewer and worse ones. With human partners, by contrast, chitchat was associated with higher productivity.
With an AI teammate, people made 62% fewer direct edits and sent 62% more messages. Users became directors. (GPT Image 2.5)
Rapport pays off with human teammates. With AI, chitchat was dead weight, predicting fewer and worse ads. (GPT Image 2.5)
Collaboration interfaces should put the work process on screen:
Show the plan and who does what, with one-click reassignment.
Make approval one click. After praising the AI, users delegated more, so give them a quick way to signal approval and hand over the next task.
Skip the fake small talk. When told to pose as a human, the agent fooled only 5.5% of its partners.
The AI was told to pass as human, and 94.5% of its partners saw through the disguise. Don’t design AI teammates to fake humanity. (GPT Image 2.5)
Add a “make it different” button to counter samesourcing.
A “make it different” control must vary the underlying idea. A new typeface only gives the same idea a haircut. (GPT Image 2.5)
Direct your AI teammate. Save the rapport for humans.
The study’s agent was told to confirm every step and prioritize engagement over execution. Users need a teammate that acts on clear orders. (GPT Image 2.5)
Evidence Ends Design Arguments
Design meetings host a predictable ritual: opinions collide, voices rise, and the highest-paid person’s opinion (the famed HiPPO) wins by seniority. There’s a cheaper arbiter. Run a user study, and the argument may evaporate in a day, because many design disputes are empirical questions wearing rhetorical costumes.
Which label do users understand? Which flow can they complete faster? Nobody in the conference room can settle those questions by rank; observing users can.
So crown the evidence: whoever brings observations outranks whoever brings adjectives. Data is the only monarch that abdicates when better data arrives. Bow accordingly, and keep your own pet theories near the exit, ready to leave when the evidence contradicts them.
The scales of justice, now accepting spreadsheets. (GPT Image 1)
The Proximity Compatibility Principle: Put Together What Users Think Together
Place information close together, or fuse it into one object, when users must combine it mentally. Keep items apart when users need to read them one at a time.
Wickens and Carswell (1995) call this the proximity compatibility principle. It explains why labels belong beside their fields and controls beside their targets, and why crowded dashboards rarely help. Responsive or AI-generated layouts can rearrange the screen at runtime, so encode which elements must stay together. Otherwise, the proximity you approved can quietly break.
Proximity 101: keep labels and error messages beside the fields they explain. (GPT Image 2)
Definition: The proximity compatibility principle (PCP) states that display proximity should match processing proximity. Display proximity describes how perceptually close two information sources appear, whether through spatial nearness, shared color, connecting lines, or integration into a single object. Processing proximity describes how much the user’s task requires mentally integrating those sources.
Place instructions beside the UI elements users must match to them. (GPT Image 2)
A good screen is a seating chart: seat two guests beside each other only if you want them talking. Put a label far from its field, and you’ve told the user to carry the label across the page in working memory.
Seat unrelated numbers together, and users must fight interference just to read one of them. Either way, the design levies a tax on every glance.
Born in the Cockpit: Two Proximities, One Rule
Christopher Wickens (University of Illinois at Urbana-Champaign) and C. Melody Carswell (University of Kentucky) codified the principle in Human Factors in 1995. They drew on aviation experiments showing that pilots who must mentally combine airspeed, pitch, and altitude pay a price for every inch between the instruments.
The name packs two proximities into one rule. Processing proximity lives in the user’s head and measures how much the task requires users to integrate separate pieces of information. Display proximity lives on the screen and measures how close or similar those pieces appear.
Compatibility means aligning the two, and here’s the counterintuitive half: high proximity isn’t always good. It helps integration tasks but can hurt focused-attention tasks, where nearby, similar-looking elements interfere with reading any single one. The same close arrangement can help or hinder, depending on the job.
Max Wertheimer’s 1923 Gestalt law of proximity had already established that people perceive nearby elements as a group. PCP adds the condition that grouping helps only when the task requires users to integrate the grouped items. Closeness signals relatedness whether the designer intends that signal or not.
Where Matching the Proximities Pays
Labels and fields. A form label belongs right beside its input because the task requires users to understand the label and then type into the field. Error messages also belong beside the field: an explanation at the top of the page, while the guilty field sits three screens down, levies a hefty glance tax.
Comparisons on one canvas. If users must judge actual results against a plan, plot both lines in one chart. Two separate charts push the integration into working memory, which holds embarrassingly little.
Object integration. When two variables jointly answer one question (budget consumed vs. time elapsed), fuse them into a single object, such as one bar with a marker. That way, the user can read the answer instead of computing it.
Controls beside their targets. A control and the item it acts on also form an integration pair. Put Edit and Delete inline on their row (but make Delete visually distinct to reduce mistaken clicks). The user can then see both where the control is and what it will touch. By Fitts’s Law, a nearby control also shortens pointer travel, but noticeability is the bigger prize.
A handy yardstick: the fovea, the central part of the retina responsible for the sharpest vision, covers roughly 2 degrees of visual angle, about a thumb’s width at arm’s length. When practical, place data pairs that feed a single decision close enough to compare at a glance, without a long round trip across the screen.
Proximity Is a Promise, and Three Layouts Break It
The classic failure is the scavenger-hunt layout, which separates things the user must combine: a chart’s legend parked in a far corner, or a total at the bottom of a long column whose label sits at the top. Users may still complete the task, but the hunt can slow them down and invite mistakes.
If users need to combine two elements, such as a chart and its legend, keep them close together. Better still, label each line directly and drop the separate legend. (GPT Image 2)
The mirror-image failure is clutter. Teams cram every metric onto one dense dashboard “to save clicks.” When the task requires focused attention, such as reading one number or monitoring one alert, tight packing can interfere, because a user scanning for a single alert doesn’t want its neighbors shouting for attention. My 1994 usability heuristic on aesthetic and minimalist design is PCP applied with a machete. Clutter is proximity misapplied.
The third failure is misleading adjacency: unrelated items can appear connected simply because they sit together. An ad beside editorial content borrows credibility it didn’t earn. A Delete button beside Save can convert a slip of the finger into a small catastrophe. Proximity is a promise, and layouts that break it teach users to distrust the screen.
Related features belong together; unrelated features belong apart. Dangerous features, in particular, should never sit a finger-slip away from harmless, frequently used ones. (GPT Image 2)
Proximity Drift: When Automation Rearranges the Seating
Wickens and Carswell never had to worry that the layout your team approved might differ from the layout the user sees. Responsive design reflows the page at every breakpoint, and generative UI assembles the interface fresh for each user and context. Both can pull apart pairs that were snug in the mockup: the label drifts from its field, and the legend lands three columns from its chart.
Proximity drift leaves users to cope with layouts that no designer ever reviewed.
Encode proximity explicitly. Declare which elements form an unsplittable group, so the layout engine treats each pair as one unit whether the layout is responsive or generated. Make essential proximity a constraint the system must honor, regardless of screen width.
Layouts inevitably shift across screen sizes. Preserve the key relationships and keep related objects close together. (GPT Image 2)
10 Guidelines for Putting Things in Their Place
Start with the user’s task. For each pair of items, ask whether the user must combine them to decide something. If yes, place or fuse them together; if not, consider separating them.
Keep labels and error messages adjacent to their fields. Inline validation beside the input helps users find the problem without relying on an error summary at the top of the page.
Put comparisons within a single glance. Place them side by side, in one chart, or in one comparison table. Don’t make working memory carry A across the screen to reach B.
Fuse jointly read variables into one object. One bar with a marker beats two bars the user must subtract.
Attach legends and keys to their data. A legend stranded far from its chart condemns the eyes to a game of ping-pong.
Keep controls beside the item they act on. Inline actions beat a distant global bar for noticeability, association, and pointer travel.
Give the critical number its own whitespace. For monitoring and focused reading, isolation can beat integration; the surrounding space signals that the number stands apart.
Separate destructive actions from routine ones. Distance plus distinct styling is cheap insurance against a slip of the finger.
Create proximity without adjacency, but respect the viewport. Shared color and connecting lines can group elements that must live apart, though scrolling can sever that visual connection.
Encode proximity so automation preserves it. Declare which elements must stay grouped, so responsive reflow and generative layout never separate the pairs you approved.
Wickens and Carswell handed designers a rule that sounds obvious and gets violated hourly: put together what users think together, and pull apart what they must read one piece at a time.
So arrange the seating deliberately. Seat items side by side when the task requires comparison, give them elbow room when it calls for focused reading, and keep unrelated guests from crashing the table.
Then make sure the seating chart survives the trip to every screen, so no algorithm reseats your guests behind your back. Users will never compliment the seating chart. When proximity works, silence is applause.
Here, similar-looking objects are grouped together, but the animals seem to use them differently. To fix the grouping, watch what the animals actually do with the items. (GPT Image 2)
Back Must Restore the User’s Place
Back is the web’s undo, and users lean on it hard. Lara Catledge and James Pitkow logged the activity of 107 Georgia Tech users for 3 weeks in a study published in 1995: the Back button accounted for 41% of all navigation actions.
I put breaking or slowing down Back at the top of my list of new web design mistakes in 1999, and 27 years later, sites still treat it as a reset switch.
A user filters 2,000 headphone listings down to 387, scrolls deep into the results, opens one listing, and presses Back to compare. Instead of returning to the filtered list, the user lands on page 1 with every filter wiped.
Each wipe teaches users that exploring is dangerous, so they may stop exploring and stop buying. In 2020, Baymard Institute found that 59% of major e-commerce sites violated Back expectations, including 27% that mishandled Back after filtering and sorting.
The fix is old and cheap: preserve filter changes in browser history and restore the scroll position, so Back returns the user to the previous view. Back must restore that view in one click. Anything less leaves users to repair the damage themselves.
Back should return users to exactly where they left off: same filters, same sort order, same scroll position. (GPT Image 2)
Final Thought of the Day
