Design Streaming AI Responses

AI Interfaces · ai, loading, motion

Updated 2026-07-28

Start showing output as soon as there is any. Streaming turns a long wait into visible progress, which is the single biggest perceived speed win an AI interface has.
Keep the scroll under the user's control. Following the output automatically is helpful until the user scrolls up to read, at which point following becomes a fight.
Do not reflow what has already been written. Text that re-wraps or re-renders as more arrives is unreadable while it is being produced.
Show the phases honestly when there is a delay before the first token, such as retrieving sources or running a tool, rather than sitting on an empty box.
Make stopping obvious and immediate. A visible stop control is essential, and it must actually halt the work rather than only hiding it.
Handle interruption cleanly. A partial answer should stay on screen, be marked as incomplete, and be regenerable without retyping the question.
Announce completion for assistive technology instead of streaming every fragment, which would be unusable, and keep the finished text reachable as a whole.
Do not let formatting arrive last. Half-rendered structure that snaps into place at the end undoes the calm that streaming created.
Keep the input available during generation so the user can prepare their next message rather than waiting for permission to think.

Streaming is a pacing device, and its purpose is to make waiting legible rather than to make it look busy.