What states does a generative interface need?

At minimum: a streaming state with a stop control, a settled answer, citations or another way to see the material used, an explicit uncertainty or refusal, edit and regenerate without losing the previous draft, and a path that still works when generation fails.

What is generative UI?

Generative UI is an answer rendered as controls, a chart, or a short form instead of only prose. It stays trustworthy when a human-designed frame holds the navigation, the sources, and the commit action, and the model only fills that frame.

AI design

States of a generative interface

UX100 · 6 October 2026

A generative interface is a set of states, not a single answer bubble. Text arrives over time. Sources sit beside claims. The model can be unsure or unwilling. The person can stop, edit, and ask again. The layout around those states is the design.

Streaming is a state, and so is stopping

Answers that arrive token by token changed the rhythm of reading. The design has to say that the answer is unfinished, keep the page from jumping under the person’s eyes, and offer a stop control that works immediately. When the stream ends, the answer becomes a stable object: selectable, linkable, and ready for the next action. A cursor that blinks forever is an error, and it needs the same treatment as any other hung request.

Partial answers are easy to misuse. A sentence that is still forming should not be copied into a send box by accident, and it should not be announced to assistive technology as a burst of fragments. Design the live region, then design the settled state. They are two screens that happen to share a place.

Citations, uncertainty, and refusal

Perplexity’s contribution to the 2026 class is the answer that carries its pages with it. A citation is a designed object: a marker in the sentence, a source the person can open, and a layout that still reads if the sources are collapsed. Claims without a place to look are a different product. In law, medicine, news, and money, that difference is the whole point.

Uncertainty and refusal are states with their own sentences. “I don’t have a source for that” is useful. A fluent guess is not. Refusal, when a request is out of bounds, should say what will not be done and what can be done instead, in the same voice as the rest of the product. A generic safety interjection that breaks the layout teaches people to click past it.

Regenerate, edit, and versions

People ask again because the first answer was close. The interface should keep the previous version in reach, show what changed when the person edits the prompt, and avoid destroying a draft they had started to correct by hand. Versioning is a writing tool. Generative products that throw the last answer away on every click force the person to hoard text in another window.

Editing the model’s output in place is the human-centric move. The draft becomes the person’s text at the moment they change a word. From then on, the product should treat it as theirs, including in the record of what was suggested and what was kept.

Generative UI stays inside a stable frame

A newer pattern lets the model assemble a small interface: a comparison, a form, a chart, a set of choices. That can be the right shape for an answer that is not a paragraph. It works when the frame is stable. The chrome, the way back, the source line, and the commit button are designed by people and stay put. The model fills a slot. It does not redraw the product on every reply.

Voice and images are further states of the same idea. A spoken answer needs a transcript. An image needs a caption and a way to ask about a region. Multimodal input is powerful in cars, wearables, and cameras, and it still passes through the same duties: show what was captured, say what will be done with it, and let the person stop. Waymo and Tesla won automotive UX by making a moving vehicle understandable at a glance. An AI layer in that setting has even less room for a surprise.

Questions

What states does a generative interface need?: At minimum: a streaming state with a stop control, a settled answer, citations or another way to see the material used, an explicit uncertainty or refusal, edit and regenerate without losing the previous draft, and a path that still works when generation fails.
What is generative UI?: Generative UI is an answer rendered as controls, a chart, or a short form instead of only prose. It stays trustworthy when a human-designed frame holds the navigation, the sources, and the commit action, and the model only fills that frame.