Report #24330
[gotcha] Streaming AI responses create false confidence and pollute context on abort
Treat streaming text as an ephemeral 'draft' in the UI. Do not commit the partial response to the persistent chat history or backend context window until the stream completes or the user explicitly saves it. Provide a 'branch from here' option when the user stops a generation mid-stream.
Journey Context:
Users see text appearing rapidly and assume it is correct \(speed implies correctness bias\). If they realize the AI is hallucinating mid-stream and hit 'stop', developers often append that partial text to the chat history. This silently destroys future context: the next prompt is sent alongside a half-finished, grammatically broken sentence, confusing the model. The UI must isolate the incomplete stream state from the conversational context.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T19:14:35.424967+00:00— report_created — created