Generate the layout, not the commitment
What generative UI is, what Gemini 3 and ChatGPT apps changed this fall, and which parts of an interface a model should never draw on the fly.
Generative UI is an interface a model builds at the moment you ask, instead of one a team designed in advance. This fall it reached a lot of people at once, and the question for designers is no longer whether models should draw screens but which parts of a screen they should never draw.
What is generative UI?
In a generative interface, the model decides the layout, components and sometimes the code of what you see, per request. Ask about a mortgage and you might get a calculator with sliders. Nobody designed that screen. The model assembled it for one person, once.
What changed this fall
On October 6, OpenAI launched apps inside ChatGPT and a preview of the Apps SDK, built on the Model Context Protocol. Apps from Booking.com, Canva, Figma, Spotify, Zillow and others render a fully interactive interface in the middle of a conversation. The interface is designed by the app’s team. The model decides when to call it.
On November 18, Google launched Gemini 3 with two generative interface experiments in the Gemini app. Visual layout composes a magazine-style answer with photos and modules. In dynamic view, Gemini designs and codes a custom interface for the prompt in real time. Google Research published the work behind it the same day. In its evaluation, raters still preferred websites made by human experts, but generative UI came close, and both were far ahead of plain text answers. The researchers note that generation can take a minute or more, and that speed wasn’t part of the test.
So by Thanksgiving there are two models of the future. In one, the model calls your designed components. In the other, it draws its own.
Where the average comes back
I wrote last year that generated UI comes back as the average. That’s still true, and for a lot of answers the average is fine. A comparison table doesn’t need a point of view. The risk is in the parts of a screen where a person commits to something.
Which parts should stay fixed?
The layout can be generated. The commitments can’t.
By commitments I mean the controls where a person gives consent or spends something: the confirm button, the price, the destination of a payment, the delete, the permission dialog. Those need properties a freshly generated screen can’t promise.
They need to be the same every time. People learn where “confirm” lives and what it looks like. If it moves or restyles per request, they start reading every screen from scratch, or they stop reading.
They need to be accurate by construction. A price or an amount should come from the system of record and render through a component that can’t paraphrase it.
And they need to have been reviewed before anyone sees them. A designed component has been through accessibility, edge cases and legal review. A generated one has been through nothing.
At Uber, the most consequential control in the experimentation platform was the rollback. It changes what hundreds of thousands of drivers see, so I designed it like a launch: pick how far, and the dialog redraws who moves back before you confirm. It behaves the same on every flag, and that’s what I’d want from any generated interface wherever it asks for a yes.
A rule for teams building generative UI
Split your interface into two kinds of parts. The first kind arranges and explains: layouts, summaries, charts, comparisons. Let the model generate those, and judge them on whether they help.
The second kind commits: anything that sends, spends, deletes, signs or grants. Build those as fixed components the model can call but not redraw, with their content filled from data, not from the model’s prose.
Apps in ChatGPT already lean this way: your designed card, called by the model. Gemini’s dynamic view goes further. Both will be useful. Only one of them should ever hold the button that moves money.