UX Roundup: Death of AI–Human Synergy | AI Answers Design Briefs | Norse vs. Greek Gods | Obviousness | Accelerating Science | Toggle Groups | Inconsistency | Seamful Design | Input Format Flexibility

Summary: Combining AI and humans produces no synergy; often AI alone is best | AI meets 99% of design-brief criteria but sticks to its house style | New video about Norse vs. Greek gods | An obvious design may be boring, but it beats a clever one | How AI accelerates scientific discovery | Toggle groups beat dropdowns for small option sets | Inconsistencies confuse users | Some UI seams give users information they need | Accept input in the user’s preferred format

UX Roundup for November 16, 2026 (GPT Image 2.6)
Human + AI < Best Solo Performer: No Synergy in 420 Studies
A self-updating meta-analysis of 420 studies finds that AI beat humans in 68% of head-to-head comparisons, and that human–AI teams lost, on average, to whichever solo performer was stronger. Test the contenders, pick the winner, and let it work alone. Humans still hold some territory, but it shrinks with every AI release.

Put a human in the AI’s loop, and the result often gets worse. At least the human’s contribution is easy to spot. (GPT Image 2.5)
Adding AI to a human makes the human better. Adding a human to an AI often makes the AI worse. That’s the verdict from AutoAnalysis, a meta-analysis by Michelle Vaccaro of Harvard Business School and co-authors from MIT that pools 1,955 effect sizes from 420 papers on people working with generative AI.
A year ago, in When Humans Add Negative Value, I argued that the “human in the loop” paradigm should die or shrink. The new data confirms that the human in the loop is often a human in the way.
Research That Updates Itself
Traditional academic publishing can’t keep up with AI. A study can languish in peer review for a year while the model it tested is superseded. A 2026 paper about a 2024 model is a history paper. Meta-analyses lag even further behind. They’re snapshots of snapshots, frozen while the field exploded from 471 papers with “human-AI” in the title in 2024 to 1,207 in 2025.

Peer review delivers its verdict at a tortoise’s pace, long after the model has been superseded. By the time the acceptance letter arrives, the AI is 4 versions newer. (GPT Image 2.5)
AutoAnalysis swaps the snapshot for a liveprint, a publication that revises itself as new evidence arrives. AI agents do the searching, screening, reading, and number-crunching. So far, they’ve sifted 15,495 records and reviewed 2,403 papers to find the 420 that qualified. For a human, that’s more than 8 months of full-time work (my estimate: 1 minute per abstract, 30 minutes per paper).

AI agents screened 15,495 records and reviewed 2,403 papers to press the effect sizes out of 420 qualifying studies. For a human researcher, that’s more than 8 months of full-time work. (GPT Image 2.5)
Better yet, the authors publish their results at auto-analysis.dev and update the site monthly. Bravo to Vaccaro and her co-authors. Medicine has experimented with living systematic reviews since 2014, but each update consumes human labor. AI makes updates cheaper, so research can come closer to keeping pace with the AI it studies.

A liveprint is never finished. AutoAnalysis rebuilds its results every month as new evidence arrives. (GPT Image 2.5)
My main quibble: by admitting only peer-reviewed papers, AutoAnalysis inherits the journals’ delays, so many of its studies probably tested since-retired models. (Validation is also thin, since the authors spot-checked only 30 papers per pipeline stage.)
AI Won 68% of the Head-to-Head Matchups
Only 171 of the 1,955 effect sizes (8.7%) came from experiments that also measured AI working alone. Most researchers compare human–AI teams against humans alone, see an improvement, and declare victory. But where AI did compete solo, it beat humans in 117 of 171 comparisons (68%). Humans won 48 (28%), and the remaining 6 comparisons ended in ties.

In head-to-head comparisons, AI beat humans 117 times, humans won 48, and 6 were ties. A rematch won’t help: the AI gets better with every release. (GPT Image 2.5)

In decision tasks with a right answer, such as diagnosing or forecasting, humans lost to AI by a large margin (g = −0.97). In creation tasks, such as writing or design, the two tied (g = +0.02).
Teams didn’t rescue the humans. On average, they lost to AI working alone (g = −0.20) and beat both solo performers in only 32% of comparisons. So the best policy is to find the stronger solo performer and let it work alone. Sometimes that’s still a human.

Human–AI teams beat both solo performers in only 32% of comparisons. Tie the partners together, and the faster one stumbles. (GPT Image 2.5)
My Two Exemptions Have Expired

Creative work used to be the last redoubt of human–AI synergy. In the generative-AI data, creation tasks show negative synergy, too. (GPT Image 2.5)
My article last year leaned on the same team’s 2024 meta-analysis of mostly pre-generative AI, which suggested two exemptions from negative human value: creative work (the last redoubt of synergy) and tasks where humans outperformed the AI. The overall pattern held (synergy g = −0.35 now vs. −0.23 then), but neither exemption survived.
Creation tasks now show negative synergy (g = −0.26). And AI help produced no reliable gain for humans who were stronger than the AI (g = +0.03), even though it lifted weaker humans by g = +0.53.

AI lifted the performance of humans who were weaker than the AI by g = 0.53, but gave no reliable gain to humans who were stronger (g = 0.03). (GPT Image 2.5)
What changed? Probably the AI. Earlier systems were weak at creative work, so humans had gaps to fill. Today’s generative AI creates about as well as the study participants, leaving fewer gaps for people to patch.

Earlier AI left gaps for humans to fill. Today’s generative AI fills them itself. (GPT Image 2.5)
How to Spot the Human Holdouts
So where do humans still win? Here are 5 guidelines, some data-driven and some speculative.
Assume AI wins until a test proves otherwise. AI won 68% of the matchups, and expertise was no shield: expert participants lost to AI alone by g = −0.80. Human superiority is now the claim that needs evidence.

Credentials aren’t data: expert participants lost to AI alone by g = −0.80. More diplomas won’t tip the scale. (GPT Image 2.5)
Run a three-arm test. Score a human alone, AI alone, and the human–AI team on at least 30 representative tasks (my rule of thumb for detecting a medium-sized difference). Skip the AI-alone arm, and negative human value stays buried in the results.

Only 8.7% of the data points came from experiments that also measured AI working alone. Without that solo AI score, a winning human–AI team can hide a human contribution that made the result worse. (GPT Image 2.5)
Ties go to the AI. Creation tasks were the only arena where humans held their own, and even there, the best they managed was a tie. Since AI is faster and cheaper, humans are the better choice only when their performance advantage outweighs their cost premium.
Apply the broken-leg test. In 1954, psychologist Paul Meehl argued that experts should override a statistical formula only when they know something decisive that it can’t see, as when a man who goes to the movies every week has just broken his leg. A 1989 review by Robyn Dawes and co-authors, including Meehl, found that experts claim too many broken legs. In a 1984 study they cite, experienced clinicians diagnosed brain dysfunction correctly 58% of the time, reached 75% when given the formula’s answer, and still trailed the formula alone at 83%.
History rhymes. Today’s broken legs are information the AI lacks, such as an unrecorded conversation or a product too new for any training data. Such an edge is informational, and it will erode as AI gains access to transcripts, email, and sensors. Superior human intelligence has nothing to do with it.

History rhymes: in a 1984 study, clinicians improved from 58% to 75% with the formula’s answer, but the formula alone scored 83%. (GPT Image 2.5)
Split the work instead of stacking a human on top. When humans and AI handled different subtasks, the team matched AI alone (g = +0.14, not statistically significant). But when both worked on the same task, the team lost to AI alone (g = −0.35). The classic human-reviews-AI setup falls into the losing category. Assign each subtask to whoever wins it.

Bolting a human onto an AI that already does the job adds a seat but subtracts value. Split the work instead. (GPT Image 2.5)

Every human review pass risks pecking another hole in work the machine had cut cleanly. (GPT Image 2.5)
Human Wins Are Temporary

Every AI release brings an upgrade. Human brains have gone 1,000 years without one. (GPT Image 2.5)
AI gets smarter with each release, while humans run the same brains as the Vikings. Thus, human wins are leases that each new AI model can shorten. Do for your workflows what this liveprint does for the literature, and rerun your three-arm test with every model generation.
My prediction: within a few years, the human holdouts will shrink to broken-leg cases, where a person knows something the AI can’t access. Human judgment is moving to AI. Human agency, deciding what’s worth doing in the first place, remains ours.

Human wins are leases, not deeds, and each new AI model shortens the lease. (GPT Image 2.5)

Judgment is moving to AI, but agency remains human: deciding what’s worth doing, and where. (GPT Image 2.5)
AI Designs Meet 99% of the Brief but Look Alike
Frontier AI now designs UI screens that meet the brief, but the designs look alike. That’s the finding of a new benchmark by Aman Rusia and colleagues at Kombai, who gave 15 frontier models and several human designers the same 160 briefs for screens such as dashboards and forms.

15 different AI models, but their designs still looked much the same. (GPT Image 2.5)
First, kudos for testing current models, such as GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash. An appendix even scores 18 more. Too many of the AI papers I review test models that were obsolete before publication. (Even this list will look quaint within months, since AI gets smarter every quarter.)

Most academic research on AI digs up fossil models instead of studying the current frontier, so the results are irrelevant by the time they’re published. (GPT Image 2.5)
The good news: every model met over 90% of the briefs’ criteria, and Claude Opus 5.5 and GPT-6 Astra hit 99%, edging out the human designers’ 98%.

Every model met more than 90% of the briefs’ criteria. Gold stars all around, for the same drawing. (GPT Image 2.5)
The bad news: the designs were interchangeable sameware. On a distinctiveness scale from 0 (identical) to 1 (completely different), two models’ designs for the same brief scored 0.59, vs. 0.76 for a human design paired with a model design.
Worse, moving the same brief to a different product domain barely changed a model’s design: 0.58 for AI vs. 0.90 for humans. GPT-6 Astra stood out most from its rivals but changed least across domains, carrying its house style from one product category to the next.

Different labs, same rose: designs from different models resembled each other far more than they resembled human designs. (GPT Image 2.5)

GPT-6 Astra stood out most from the other models but changed least across domains, like an architect who designs a hospital, a school, and a prison in the same house style. (GPT Image 2.5)
Ask AI for 10 Designs, Not 1

The top models met 99% of the briefs’ criteria, but their designs lacked variety. For the second part of this news item, I asked the image model to vary its style instead of sticking to the flat look from the first part. (GPT Image 2.5)
Regrettably, the study measured only AI’s first idea. Each design came from a fresh context, with no request for alternatives. (The authors acknowledge this limitation.) No good designer stops at one design, and no good AI user should. Had the researchers asked for 10 layouts built on different organizing principles, each new design could have steered away from the earlier ones. That would likely have broadened the range of solutions.

Every design in the study came from a fresh context, so for the model, every design was its first. (GPT Image 2.5)

Asking for many different alternatives is how superusers prompt. (GPT Image 2.5)
The comparison also favored the humans, since several designers could draw on inspiration libraries such as Behance, while each model worked alone.

The human designers could browse inspiration libraries. The models worked alone. (GPT Image 2.5)
Kombai sells AI design tools, so the authors have a stake in showing that general-purpose models need help. Still, the methodology looks sound.

Vendor research always deserves some skepticism, even when (as in this case) the methodology holds up. (GPT Image 2.5)
In 1996, Jan Maurits Faber and I found that parallel design (several designers working independently, then merging their best ideas) improved measured usability by 70%, compared with 18% for traditional iterative design. Back then, it took a team. Now, even with Claude Fable 5.1, the priciest model tested ($5.33 per design), 10 alternatives cost about $53.

10 alternatives from the priciest model cost about $53. The real cost is the time you need to review them. (GPT Image 2.5)

In 1996, parallel design took a team and a week. Now AI produces alternatives faster than you can explain the method. (GPT Image 2.5)
So use diamond prompting: diverge with many deliberately different designs, then converge by merging the best ideas. AI’s first design is its most average design. Ideation is free, so there’s no excuse for shipping sameware.

Diamond prompting in two breaths: diverge with many deliberately different designs, then converge by merging the best. (GPT Image 2.5)

Ideation is free with AI, but you have to ask. Most users stop at the first idea. (GPT Image 2.5)

The first design is the model’s safest guess. Ask for the ones it wouldn’t offer unprompted. (GPT Image 2.5)
New Video: Norse vs. Greek Gods
I made a new video in which the Norse and Greek gods debate which pantheon is best (YouTube). The story itself runs 5 minutes, but the video is 10 minutes long because it includes a second run-through of the story in several alternative animation styles.

New video: The Viking and Greek gods go head-to-head. (GPT Image 2.6)
I used a different workflow for this video because it’s a piece of fiction. Odin never met Zeus in the actual myths, so I didn’t want to rely on agentic video editing to stitch together the pieces I borrowed from the old stories.
I did use agentic production to have the AI shoulder the drudgery of uploading the correct reference images and prompts for each clip, including retakes of specific shots that came out wrong.

My lead characters are the two kings of the gods, Odin and Zeus, though Thor, Aphrodite, and Thor’s magic goats also have prominent parts. (GPT Image 2.6)
But I did the editing by hand in good old CapCut, combining the takes and vetting the agent’s generations. The agent caught several problems on its own, but it missed others that surfaced only when I watched rough cuts of my edits.
I first generated 38 different animation styles and produced 30-second clips for most of them to see how the gods would move and sound in each style. Here are 4 of them:

(Style concept images made with GPT Image 2.5)
I chose the bottom style, Paper-Cut Collage, for my video because I felt anime had been done to death and the more realistic styles were a poor match for my tongue-in-cheek handling of the myths. If you watch my YouTube video to the end, you can see the other styles in motion and judge for yourself.
After deciding on the visual style, I generated complete character reference sheets for all 4 gods, as well as Thor’s goats and chariot. I also made several location reference plates, plus reference images for the birds and for props such as Odin’s drinking horn and Aphrodite’s golden apple.
These reference images (made with GPT Image 2.5 and 2.6) kept the look consistent across 38 video generations with Wan 3.0. (For photorealistic video styles, I prefer Seedance 2.5, but Wan is better suited to the stylized illustration I used for this video.)

Character reference sheets for Odin and Zeus. (GPT Image 2.5)
Despite several retakes, the final video contains some continuity errors. For example, Odin’s raven hops between his left and right shoulders from shot to shot. As always, fixing the last few percent would be half the work, so I skipped it. Just pretend that the bird flutters about whenever the camera looks away.

I didn’t use a photorealistic style in the actual video, but it makes for cool movie posters. (GPT Image 2.6)
I also had audio problems. The Viking gods were supposed to have Scandinavian accents, but Wan made them speak standard English. Experimenting with ElevenLabs v4 gave me slightly better voices, but still no true Scandinavian accents. (I should know: I have one!)
Wan supposedly supports uploading a voice track and lip-syncing the video to it, but this worked poorly in my testing, so I stayed with the voices produced natively during video generation.

Spoiler: I didn’t want a conventional slugfest, so the gods compete in other ways, including Viking mead vs. Greek wine. (GPT Image 2.6)
Watch the video on YouTube.
Obvious Beats Clever
A well-designed door needs no sign: the flat plate says push, the handle says pull. A door that needs a Push label has lost a small battle, yet it still wins the war against the alternative, a contraption of tangled pipes and competing commands that users must reverse-engineer before daring to act. My good friend Steve Krug compressed our whole discipline into the title of his best-selling 2000 book, Don’t Make Me Think.
Teams ship puzzles because of the curse of knowledge. After months inside a design, nobody on the team can see it with fresh eyes, so every internal review certifies the obviousness of things that aren’t obvious at all.
The antidote costs 5 minutes. Show the screen to a stranger for 5 seconds, then ask what the page is for and what he or she would click first. If the stranger hesitates, redesign. Obvious beats clever whenever the two compete.

Clever design gets admired; obvious design gets used. (GPT Image 2)
Galileo to Starlink Took 400 Years; AI Science Will Pay Off Faster
Google and Anthropic data show scientists using AI mainly for data analysis and writing, saving almost 7 hours weekly while untested hypotheses pile up. Scientists underuse AI for ideation, though it could also help rank those hypotheses. Like Galileo’s astronomy, AI science will pay off in unpredictable ways, but far sooner.
Science has accelerated ever since Francis Bacon, Galileo Galilei, René Descartes, and Sir Isaac Newton turned discovery into a repeatable method in the 1600s. Galileo’s telescope extended the eye. AI is a forklift for the mind, and two new reports show how scientists use it.
Scientists = AI Superusers
Mihai Codreanu and co-authors from Google, Google DeepMind, and MIT FutureTech analyzed 15 million Gemini interactions and 2,690 specialized AI models, and surveyed 637 scientists. Scientists use Gemini 2.7 times as much as their share of US jobs predicts, 47% use AI daily, and the average respondent saves almost 7 hours per week.
Johannes Hermle and co-authors at the Anthropic Institute classified 207,049 research sessions on Claude and Claude Code. Data analysis took 48% of Claude Code sessions but only 28% of researchers’ time; writing took 32% of Claude sessions, double its share of researchers’ time. Tasks in the top 20% of AI capability got 2–4 times the use their share of researchers’ time predicts. Scientists use AI where it works.
OpenAI’s January report (PDF) counted 1.3 million weekly ChatGPT users working on advanced science and math. Surely more now.
Caveat: these are all vendor studies, and their surveys probably skew toward enthusiasts. Still, the rival datasets agree.
Google also found that LLMs and specialized models like AlphaFold divide the labor. Bacon foresaw this. Salomon’s House, the research institute in his New Atlantis, employed “Merchants of Light” to gather knowledge, “Pioneers or Miners” to run experiments, and “Interpreters of Nature” to derive axioms. Google’s vision of an LLM orchestrator calling specialized models is Salomon’s House in silicon.
AI Makes Ideas Cheap, but Atoms Stay Expensive
In 1637, Descartes advised dividing each difficulty into as many parts as possible. Both reports split science into thousands of tasks. Speed up some, and the rest become the bottleneck. In Google’s survey, 44% of scientists say their main constraint has moved downstream into the lab, clinic, or field, and 41% report a growing backlog of untested hypotheses. So AI-generated ideas pile up in a labjam.
Anthropic estimates that current AI, widely adopted, could double research output in statistics. If AI did all non-physical research work, computer science and statistics could produce 15–16 times as much, but biology and chemistry only 3–4 times as much, because physical work eats about 40% of their time. I think that’s too pessimistic. At Pacific Northwest National Laboratory, automated labs already run 5–10 times as many experiments as would be practical by hand.
Verification is the other drain: 46% of scientists who save time with AI spend more than 1/4 of the savings checking its output. The Royal Society, founded in 1660 in Bacon’s spirit, chose the motto Nullius in verba: take nobody’s word for it. AI included. But AI can check, too. Claude turned Andrew Wiles’s proof of Fermat’s Last Theorem into machine-checked code in 11 days, a job expected to take years.
Safe Questions vs. Moonshots
Does AI make scientists bolder? Not yet. Anthropic’s respondents rate ideation as their most important activity, yet it accounts for only 4.2% of Claude sessions. In Google’s survey, 49% say AI steers them toward safer questions, while 28% say riskier ones. A Nature study of 41 million papers by Qianyue Hao and co-authors agrees: AI users publish 3 times as many papers, but the range of topics that science explores shrinks.
That’s unfortunate. True, empirical testing is the bottleneck, so more ideas alone would only lengthen the labjam. But AI can also winnow and rank hypotheses, so researchers test only the most promising ones at the bench. (Anthropic’s authors note that better hypotheses would pay off most where testing is costly.) Fully virtual labs that replace physical reality remain rare outside mathematics, where Lean checks proofs.
UX offers a parallel. I’ve long warned against synthetic users as a substitute for research with real customers, while applauding AI that makes empirical research faster and cheaper. For example, rehearsing interviews with AI-simulated participants led 17 of 21 researchers to revise their interview guides before meeting real users. Use AI to sharpen experiments, not to skip them.
Frontier models are bolder. In May, an OpenAI model disproved an Erdős conjecture from 1946. In August, a Claude research model raised the proven share of Riemann zeta zeros on the critical line from 42% to 67%, while its non-mathematician prompter mostly typed encouragement like “keep going.” (Management by pep talk works on AI, too.) In September, OpenAI claimed that about 10,000 agents had resolved the Navier–Stokes Millennium Prize Problem in a mere 88 hours. (Mathematicians are still checking.) Those equations are Newton’s second law applied to fluids.
Galileo Couldn’t Imagine Starlink
Medicine and drug discovery look like AI science’s biggest prizes, yet Anthropic expects the bench sciences behind new drugs to be among the slowest to accelerate. Should we despair? No. Science pays off in unexpected places.
In 1610, Galileo discovered Jupiter’s moons, then spent years trying to sell Spain and the Dutch Republic on using them as a clock for finding longitude at sea. The scheme flopped. (Try steadying a telescope on a rolling deck.) Newton explained why moons orbit and imagined a cannonball fired so fast that it falls around the Earth forever: the first artificial satellite, on paper. On June 22, 2026, United flight 14 flew from Newark to London with free Starlink Wi-Fi, beamed via satellites that obey Newton’s laws in Descartes’s coordinates. Galileo couldn’t have envisioned browsing the Web while hurtling across the Atlantic in a metal tube.
AI science will also bring benefits nobody can predict, me included. But they’ll come much faster because the fields poised to accelerate most include computer science, which builds AI itself, closing a feedback loop. I expect superintelligence by 2030.
Conclusion: The New Instrument Needs Good UX
Bacon titled his 1620 book Novum Organum: the new instrument. AI is the newest instrument, and its design matters. Claude Science, which integrates research databases, draws twice the ideation share of general Claude. Design AI tools to generate and rank hypotheses and to verify results. Fund lab robots, too, or hypotheses will rot in the labjam. Galileo’s moons needed 416 years to reach your airplane seat. AI science won’t take that long.
Alice and Zimo explain the acceleration of science in Blue-Gray Sci-Fi comic style (GPT Image 2.6):




















Toggle Groups Beat Dropdowns for 2–5 Options
A toggle group shows every option and the current selection at a glance, and it switches state with a single click. That makes it the best control for a small set of peer options. But it becomes a guessing game whenever the selected state is weakly marked or the number of options climbs past 5.
Definition: A toggle group (AKA segmented control or toggle button group) is a row of connected buttons representing a small set of options. In the single-select version, exactly 1 option is active at a time, like radio buttons (e.g., Left, Center, and Right paragraph alignment). In the multi-select version, each button switches independently, like checkboxes (e.g., Bold, Italic, and Underline).

One segment glowing, 4 resting. The entire promise of a toggle group is that the current state and all alternatives are readable at a glance, with no clicking required. (GPT Image 2)

A simple toggle switch turns a single option on or off. Easy, but limited. A toggle group extends the idea, letting users choose among more states. (GPT Image 2)
From Rigging Pins to Car Radios
The name is roughly 2 centuries older than computing. Sailors used a toggle, a pin pushed through a loop of rope, as a quick-release fastener. The electrical toggle switch (a lever flipped between two positions) inherited the word. The grouping half of the pattern has two ancestors. The first is the mechanical car radio of the 1930s: press one preset button, and the previously pressed button pops back out. Those interlocked buttons gave the GUI its radio buttons, name and mutual exclusivity included.

Selecting exactly one of many options is an old design problem, and the solutions have barely changed across the ages. (GPT Image 2)
The second ancestor is the word-processor toolbar, where Bold, Italic, and Underline have worked as independent press-in toggles for 40 years. Apple fused both ideas into the segmented control on the Mac, and the iPhone (2007) made it a daily sight. Google’s Material Design calls the same control segmented buttons.
Google’s guidance is specific: use 2–5 segments, and beyond 5, switch to a different component. My experience agrees with the ceiling and suggests 2–4 as the sweet spot. (Google’s newest Material revision renames the control yet again, to the connected button group, proving that the pattern outlives its labels. Come on, Google, enough with the vocabulary inflation. Keep the name people already know.)
Everything Visible, Everything 1 Click Away
A dropdown holding the same options shows the user the current selection but hides all alternatives behind a click. The toggle group therefore wins on the speed of both reading and acting, at a price paid in pixels:

For controls users hit repeatedly, such as a calendar switcher for Day, Week, and Month, the saved clicks compound all day long. Decision speed benefits, too. Hick’s Law (William Hick, 1952) says decision time grows with the number of alternatives, so a set of 5 or fewer visible, labeled options keeps both scanning and choosing quick. Thus, the control fits view switching (list vs. grid, map vs. satellite), small filters, and settings with instantly applied peer options, such as monthly vs. annual billing.

For tasks where users frequently switch views, the single-click activation of a toggle group shines. Saving one click doesn’t sound like much, but repeat it all day, and the spared aggravation adds up. (GPT Image 2)
Beware the Coin-Flip Control
Everything above rests on one fragile assumption: users can tell which segment is selected. Break it, and the control breaks. The result is a coin-flip control: a 2-segment toggle whose selected-state styling is so weak that reading it is a 50/50 guess. Is the highlighted segment the current state, or the button I’m supposed to press? Users can’t tell, and on a billing page where the segments read Monthly and Annual, the coin flip costs real money and earns you furious customers.

With only two options, a color difference alone can’t show which one is selected. (GPT Image 2)
The fix is redundancy. Material Design 3 adds a checkmark to the selected segment because a fill color alone can be ambiguous. And color alone fails color-blind users, roughly 8% of men. Combine a strong fill with a checkmark and clear contrast, so that the selection survives a squint test at arm’s length.
4 more ways designers sabotage the pattern:
Option inflation. At 6+ options, labels truncate, tap targets shrink, and rows wrap. Use a dropdown or a radio list instead.

Toggle groups buckle once the number of options grows. (GPT Image 2)
Icon-only mystery meat. My old friend Vincent Flanders coined the term “mystery meat navigation” in the late 1990s for controls that reveal their meaning only when poked. Icon-only toggle groups qualify unless the icons are as conventional as B, I, and U. The term is old, but the usability problem is still with us.

Unlabeled toggle group options easily become mystery meat navigation unless the icons are universally understood. Just add a label. (GPT Image 2)
Actions smuggled into states. A Delete button lurking inside a view switcher will get pressed by somebody expecting to change views. States and actions should never share a group.

Use toggle groups to switch (!) between states, and leave actions to regular buttons. (GPT Image 2)
Selection-mode ambiguity. Users can’t predict whether pressing one segment releases the others. Visual convention helps. Connected segments on a shared track suggest picking one option, while checkmarks appearing on several segments announce multi-select.
9 Design Guidelines for Toggle Groups
Cap the set at 5 options, and prefer 2–4. If you have more options, switch controls instead of stretching the row.
Mark selection redundantly: distinct fill plus a checkmark or equivalent second cue, never color alone.
Prefer short text labels of 1–2 words in parallel grammar. Reserve icon-only segments for conventions users already know, and give even those tooltips.
Never truncate labels. If labels don’t fit, you have too many options or labels that are too long. Both are problems for you to solve before users ever see them.
Apply changes instantly. A toggle group with an Apply button is a radio group in disguise; when switching is expensive, use actual radio buttons plus a confirmation.
Keep 1 segment always selected in single-select groups; a view switcher with nothing active leaves the interface in an unnamed state.
Exclude actions. Deleting, saving, and sharing are verbs; a toggle group holds states.
Signal the flavor. Make single-select and multi-select groups visually distinguishable so users can predict what a press will do.
Support the keyboard and screen readers: use appropriate group semantics, keyboard movement between segments, and announced state changes.
A toggle group exists to pass one test: can a user, at a glance and without touching anything, name the current state and every alternative? Pass it, and you’ve outclassed the dropdown. Fail it, and you’ve built a coin-flip control that squanders the glanceability you promised. If reading a control requires clicking it, the control has already failed.
Every Page Should Obey the Same Physics
Three pages, three sets of door hardware: a knob, a pull handle, and a push bar wearing a Push sign like an apology. Don Norman made this door famous in The Design of Everyday Things. When hardware needs a label, the design has flunked. Websites fail the same way, one threshold at a time. Change the navigation between pages, and users must relearn your interface mid-task, burning attention they’d rather spend on your content. Unmotivated change is pure cost. If a redesign has no user benefit, it has no excuse. Pick one physics. Keep it.

Page 3 needed instructions. Page 4 ships with a manual. (GPT Image 2)
Seamful Design: When Hiding the Seams Hurts Users
“Seamless” is the most overused compliment in tech, but hiding every system boundary also hides information users need: coverage gaps, sync state, sensor accuracy, and AI uncertainty. Seamful design deliberately exposes selected seams so users can understand, predict, and even exploit the system. Show the seams that carry useful information; hide the rest.

Kintsugi repairs broken pottery with gold, turning every crack into part of the design. Interfaces can treat their seams the same way, gilding the joints that tell users something. (GPT Image 2)
Definition: A seam is a boundary or discontinuity between components, infrastructures, or modes: the edge of Wi-Fi coverage, the handoff from GPS to cell-tower positioning, the line between cached and live data, or the switch between automated and manual control. Seamless design hides these joints. Seamful design selectively reveals them, so users can reason about what the system knows, how sure it is, and what will happen next.

It’s better to level with users about what they need to know than to hide everything under a smooth, seamless surface. (GPT Image 2)
Beautiful Seams: The Founder’s Heresy
The name has a delicious origin. Mark Weiser of Xerox PARC, the father of ubiquitous computing and the great apostle of invisible technology, turned on his own gospel. In invited talks at UIST in 1994 and USENIX in 1995, he argued that “seamlessness” was a misleading goal, because making everything seamless means reducing every tool to bland sameness. He proposed the opposite ideal: seamful systems with “beautiful seams.”
Matthew Chalmers and Ian MacColl of the University of Glasgow formalized seamful design in a 2003 workshop paper, arguing that seams such as sensing uncertainty are inevitable in mobile systems and that users can perceive and appropriate them for their own ends. (For a readable tour of the whole debate, see Sarah Inman and David Ribes, “Beautiful Seams,” CHI 2019.)
Chalmers’s group even put the idea to the test in a game. In Treasure, presented at UbiComp 2005, players roamed outdoors with PDAs, collecting virtual coins that lay outside Wi-Fi coverage and banking them by returning inside coverage, while their devices mapped signal strength as they moved. Players quickly devised tactics around the coverage edge, ambushing rivals in dead zones. The seam was the fun.
Telltale Seams in Everyday Interfaces
You already rely on seamful design daily, probably without knowing its name. The blue location dot on a map app sits inside a translucent accuracy halo, a frank admission that the system knows your position only within, say, 30 meters. That halo can keep you from turning two streets early.
Sync indicators (“Saved 2 minutes ago,” “Offline: changes will upload when you reconnect”) expose the seam between local and cloud storage. Battery percentages and signal bars are seams between you and physics. Each is a telltale seam: a revealed boundary that carries information the user can act on, by waiting, retrying, relocating, or double-checking.

Disclosing uncertainty in a location UI helps users match the screen to the street. (GPT Image 2)
AI has made this old idea urgent again. A model that shows its sources, flags low confidence, and announces which tool an agent just invoked is displaying telltale seams. Users can then calibrate their trust. A model that answers every question in the same smooth, confident voice, right or wrong, offers seamlessness of the worst kind: the hallucination arrives wearing the same suit as the verified fact. The cure is my first usability heuristic from 1994 (visibility of system status), tailored for 2026.
Two Failure Modes: Fake Seamlessness and Seam Spaghetti
Designers can err in either direction, and both failures hurt.
Fake seamlessness hides joints that users need to see. Consider the cloud drive that silently fails to sync, letting someone edit a stale file all afternoon; the assistant that switches from retrieved facts to free-form guessing mid-answer without any signal; or the “unlimited” streaming service that slips into low resolution unannounced.
When the hidden seam finally ruptures, the user pays the whole bill at once, plus interest in lost trust. Silence about system state is concealment dressed up as polish.

Don’t hide information users need, such as rate limits. Better to let users decide when to save capacity than to have the system fail unexpectedly because the “unlimited” offer turned out to be a lie. (GPT Image 2)
Seam spaghetti is the opposite sin: exposing every pipe and rivet. Think of raw stack traces in user-facing dialogs, “Error 0x80070057,” and toast notifications about the health of internal microservices. These are seams no user can act on, presented in a language no user speaks. Showing them merely dumps the developer’s debugging costs on the customer.

Don’t overdose on seamful ideology. Most seams should stay hidden, since exposing them only adds noise. (GPT Image 2)
A single question steers you clear of both failures: can the user do something useful with this boundary? If yes (retry, wait, move, verify, switch modes, or lower expectations), reveal it in task language and pair it with the next step. If no, log it for your engineers and keep the interface calm. And when you do reveal a seam, make it quiet, glanceable, and informative. Gold in the cracks, not rubble on the floor.
6 Design Guidelines for Beautiful Seams
Reveal a seam whenever its state can change without user action. Connectivity, sync status, sensor accuracy, and data freshness all shift on their own, so users need a visible indicator to plan around them.
Apply the actionability test before exposing anything. If knowing about the boundary lets a user retry, wait, relocate, verify, or switch modes, show it; if not, it belongs in the logs.
Translate every revealed seam into task language with a next step. “You’re offline; edits will upload when you reconnect” helps, whereas an error code merely transfers anxiety.
Show uncertainty visually where it affects decisions: accuracy halos on maps, confidence labels on predictions, and “last updated” timestamps on any data that ages.
Give AI features provenance seams. Display sources, cite retrieved material, announce tool use, and mark clearly when the system is extrapolating instead of reporting, so users can calibrate their trust rather than having it manufactured for them.
Never let a failure be silent. A visible “not saved” warning beats an invisible loss every time; a system that conceals its own breakdowns is user-hostile, however smooth it looks in the demo.
Be Liberal in What Your Forms Accept
Jon Postel wrote the robustness principle into the TCP specification in 1980: be conservative in what you do, be liberal in what you accept from others. The Internet runs on it, yet the forms built on top of the Internet do the opposite. A phone number with parentheses gets bounced, then the same digits with spaces, then the bare digits, because the field secretly wanted dashes. Dates and postal codes get the same treatment. Every one of these entries carried the correct information, and the computer, which can strip punctuation a million times a second, asked the human to do it by hand.
The rule: if the underlying information is correct, accept it and normalize it yourself. Strip spaces, dashes, and parentheses from phone numbers. Parse any unambiguous date. Show an example next to the field, and when input is beyond rescue, say exactly what’s needed. Users came to give you their data, and they did. Take it.

Wrong format. Right information. Guess which one counts. (GPT Image 2)
Final Thought of the Day




