TL;DR — Not every AI interaction should be a chatbot. Real work sorts into three interface shapes: Direct GUI (point and click), Headless Agent (runs in the background), and Generative UI (builds a bespoke screen per question). The skill is matching the shape to the task. This note includes an interactive demo and a 3×3 decision matrix.
Field Note · Visual Interface
The app that builds itself for one question.
Chat gave every problem the same little text box. Generative UI does something stranger and more useful — it assembles a bespoke, throwaway interface around the exact thing you asked, then forgets it when you're done.
For three years the answer to "how should people use AI?" has been a single, stubborn shape: a chat box. You type, it types back — a wall of markdown you then have to read, parse, and act on yourself. It was a huge leap over a blank prompt. It is also, for a surprising amount of real work, the wrong tool.
Because not every question wants a paragraph. Ask a model to help you dress for a winter wedding and a wall of text is a downgrade from a page you could actually look at. Ask it to explain fractals to a curious ten-year-old and the right answer is something they can drag a slider on. Ask it to compare database options for your workload and you want a dynamic simulator, not prose.
That is the bet behind generative UI: instead of returning content, the model generates the interface — a small, custom, interactive application built for one prompt and thrown away after. Google's research team put it plainly: not just content, but "immersive visual experiences and interactive interfaces… automatically designed and fully customized in response to any question, instruction, or prompt."
Direct GUI
The user points at and manipulates known objects. Best when the action is small and the context is clear.
Editing, approving, selecting, scanning.
Stitching together multiple systems/policies.
If they can say "change this line," give them the line.
Headless Agent
The work runs in the background. Best when tasks are long, multi-step, and procedural.
Running tool chains, tracing data, retrying steps.
Comparing tradeoffs or steering paths visually.
If the job is "do the analysis, show errors," hide the screen.
Generative UI
The layout adapts to the question. Best when teams need to see planning constraints dynamically.
Temporary surfaces for messy planning/tradeoffs.
Rebuilding layouts adds reading time instead of saving it.
If each team needs a different lens on the same plan, generate it.
Watch it happen
Here is the same prompt bar, three different questions. A chatbot would answer all three with text. Generative UI answers each with a different tool. Pick a question and watch the screen assemble itself.
↑ Choose a prompt. The same box will build a different interface for each one.
Why this is different, not just prettier
It is tempting to file generative UI under "nicer formatting." It is more than that. A chat answer is static — the model's one shot at being useful, frozen as text. A generated interface is a place you can keep working: adjust the fractal's depth, swap a jacket, get the next problem. The output stops being a message and starts being a small machine.
When Google ran the comparison, humans strongly preferred generated interfaces over the standard markdown reply — a "substantial gap" — landing second only to sites hand-built by human experts. The honest catch is speed: generating a real interface can take a minute or more, and that minute is not free. Which is exactly why generative UI is a choice, not a default.
It's one of three moves, not the only one
The mistake would be to make everything generative, the same way we made everything a chatbot. Most real work sorts into three interface shapes — and the skill is matching the shape to the task in front of the team.
Chat was never the destination. It was the placeholder we all agreed to while we figured out what the interface should have been.
We must distinguish between three distinct product interfaces, choosing the correct shape for the cognitive overhead of the target workload.