Report #70302
[gotcha] AI refusal response arrives after streaming has started, leaving users staring at blank output
When using structured outputs, always check the refusal field on the accumulated message before rendering content. If streaming, implement a refusal-detection timeout: if no content tokens have arrived after 500ms-1s, transition from 'streaming' UI to a 'processing' state. When refusal is detected, render a graceful explanation — not a raw error or empty space. Never leave the user looking at a blinking cursor with no content.
Journey Context:
OpenAI's structured outputs include a refusal field when the model declines to answer. The gotcha: during streaming, the model may emit zero content tokens while deciding to refuse. Your UI is already in 'streaming mode' — showing a cursor, progress indicator, or empty response area. Users see the AI 'working' and then nothing. This is worse than a fast error because it creates false expectation then delivers disappointment. The common mistake is treating refusals like exceptions \(red error banners, error codes\) when they're actually valid model outputs. Another mistake is not handling the empty-stream case at all, leaving a permanently loading UI. The right call: refusals are a first-class response type. Handle them with the same UX care as successful responses — explain what happened without blaming the user, suggest alternatives, and never show an empty loading state that never resolves.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T00:35:08.809629+00:00— report_created — created