Predictable Generative UI

Early in my time leading design teams at scale, a team adjacent to mine rounded the corners of listing cards by four pixels, on a surface serving hundreds of millions of people. It caused a 10% regression. The change wasn't wrong in isolation; it was wrong at scale, inside an interface people had already learned. That experiment wasn't about predictability. It taught me something adjacent and more fundamental: once people have learned a surface, even tiny changes to it carry enormous consequences.

Hold that thought against what the industry is building next: interfaces that regenerate themselves (layout, controls, hierarchy) every time you open them. I've believed for a long time that a system earns trust by being predictable in how it behaves. Generative UI is the hardest test that belief has ever faced.

So I went back through thirty years of adaptive-interface research and the last two years of generative-UI evidence, and I wrote down what I found. I wrote it as a specification, not an essay, because this moment doesn't need another opinion about AI. It needs constraints.

Nielsen Norman Group defines generative UI as an interface “dynamically generated in real time by AI to fit the user's needs and context.” Not one design for a billion people: a design for you, for this task, for this moment. Designers stop crafting screens and start defining outcomes.

The industry has committed real money to this. Google generates entire interfaces per prompt in Gemini. OpenAI and Anthropic ship interactive apps inside their assistants. Vercel and Thesys sell the infrastructure. A €10M German consortium (Audi, MAN, Valeo) is building it into vehicle interiors.

The cleanest evidence comes from Gajos, Everitt, Tan, Czerwinski, and Weld at CHI 2008. They isolated the two properties everyone argues about: predictability (can you anticipate what the interface will do) and accuracy (does the adaptation actually match what you need).

Both improved satisfaction. But accuracy had the stronger effect on performance and utilization, stronger than the authors themselves expected. Read carefully, that's not a license to generate freely. It's a warning with two edges: when the system adapts correctly, users tolerate some surprise. When it adapts wrongly (and every generative system sometimes will), unpredictability makes the failure doubly punishing. You guessed wrong, and I can't find anything.

None of this is users being stubborn. It's accounting. Kahneman and Tversky showed that losses loom roughly twice as large as equivalent gains, and Samuelson and Zeckhauser showed we defend the current state even against objectively better alternatives. A learned layout is a possession; Kahneman, Knetsch, and Thaler called it the endowment effect. Kim and Kankanhalli measured the same force in enterprise software: users resist beneficial systems because switching costs are paid now and benefits are paid later.

Run the math for a generated layout. A rearrangement that is 10% more efficient still lands as a loss, because the user pays the reorientation cost immediately and collects the efficiency slowly, against a roughly two-to-one penalty on the loss side.

Skilled interface use isn't thinking; it's procedural memory. Cohen and Squire showed it's a separate memory system from knowing facts, and Karni's fMRI work showed practice physically rewires motor cortex. After enough repetitions, reaching for a control bypasses conscious search entirely. That's why you can type without looking, and why Fitts blamed “human error” on design fifty years before we shipped software.

The spatial-memory literature sharpens it: Scarr, Cockburn, and Gutwin showed stable layouts let experts retrieve commands from remembered location, skipping visual search; the entire novice-to-expert transition depends on positions holding still. Labels, colors, and content are cheap to change. Positions are expensive.

The evidence doesn't say never adapt. It says: change what the interface shows. Don't change where the interface lives. Watch the same generative power spent two ways:

Define a spatial grammar of fixed zones (nav, content, actions, status) that the model cannot re-decide. Generative freedom lives inside the zones. The header stays a header. The primary action stays where the hand learned it. This is Nielsen's “stable skeleton”; it's also how every platform that actually shipped in 2026 chose to work.

Stage the rollout. New elements arrive visibly differentiated; removed ones collapse rather than vanish; the user can always trace the delta from last time. Findlater's group showed the mechanism works: gradual-onset “ephemeral” menus beat static ones when predictions were good, without the disorientation of reordering. And when YouTube replaced its entire design in 2017, it shipped the new look as an opt-in preview with revert for months before making it default. The biggest sites already treat radical change as a staged negotiation. Your model should too, remembering from §3.3 that “gradual” means signaled, not snuck.

Gajos made accuracy the strongest lever we have: getting the right controls in front of the user beats getting interesting ones. The 2026 evidence adds a twist: Romero's team found AI-generated interfaces score well on usability and poorly on originality. Models generate conventional layouts. So the real predictability threat isn't wild design. It's variance: a systematic review of generative no-code tools found identical prompts producing substantially different interfaces across runs of the same tool. A generated checkout still needs cart, shipping, payment, confirmation, in that order, every time it's yours.

Remember the hidden cost from §3.3: the more accurately an interface adapts to your current task, the less of the system you're aware of. An AI optimizing for your immediate intent is an AI quietly deciding what you'll never discover. The remedy is structural: a stable “show everything” escape hatch, in a fixed location, that restores the comprehensive view, plus an undo for any layout the model chose for you. The endowment effect says a learned interface is a possession; revert is how the user keeps the title.

An unexplained adaptation feels like a colleague rearranging your desk over lunch. The fix costs one line: “Showing editing tools because you're cropping.” Festinger's cognitive-dissonance work explains why this matters more than it looks: when reality contradicts “I know how this works,” people resolve the conflict by rejecting the tool, not by relearning it. An annotation converts the change from a violation of the user's model into an update to it. This is also where my own conviction sits: a system earns trust by being predictable in how it behaves, and explicit about what it doesn't know.

Jakob's Law: users spend most of their time in other interfaces, so they arrive carrying a composite model of how software works. A generated interface doesn't get to opt out of that model. It inherits it. Nielsen put a number on when deviation is justified: only when the alternative measures at least 100% higher usability. Not ten percent better. Twice as good. Swipe-back, pull-to-refresh, how a toggle behaves, what a modal does: these are immutable rails. Encode them as constraints the generator physically cannot cross, the way OpenAI's and Anthropic's app frameworks already do with reviewed component vocabularies.

Look at what the platforms actually built, versus what the demos promised. OpenAI renders developer-authored, reviewed components; the model chooses which, never what. Anthropic's MCP Apps run designed UIs in sandboxes. Vercel's AI SDK moved from freeform streaming toward typed tool-to-component mappings. Thesys generates a constrained spec into a fixed design system. In the academy, Cao and Xia's CHI 2025 system generates a task data model first and derives the interface by rule: consistency by construction.

Every one of these is REQ‑01 and REQ‑07, implemented independently, by teams that mostly haven't read the adaptive-interface literature. They rediscovered it the expensive way. The remaining six requirements are where the next generation of these systems will be won. They're cheap to adopt now, and brutal to retrofit after your users have learned three different layouts.

One more thing, about the document you just read. The title block never moved. The clause numbers told you where you were. Every sheet changed completely (evidence tables, live figures, normative fields) inside a structure that never did. You read twenty-two screens of dense material and never once wondered where you were.

That's the whole argument.