Report #74339
[gotcha] Streaming text creates false conversational turn-taking expectations — users think they can redirect the AI mid-stream
Clearly signal that the AI is generating a complete response, not engaging in real-time dialogue. During streaming: visually de-emphasize or disable the input field, show a 'generating response...' state, and use a distinct visual treatment \(e.g., different background, pulsing border\) for the streaming area. After the response completes, re-enable input with a 'follow up' affordance. Never accept mid-stream input that appears to redirect the AI but actually queues for the next turn without clear communication.
Journey Context:
When users see text streaming in character-by-character, their brain maps it to human conversation — they expect to be able to interrupt, redirect, or interject, like they would with a person speaking. But LLMs generate responses as a single forward pass \(or stream tokens as they're generated with no ability to change course based on new input\). If your UI allows typing during streaming, users think they're having a real-time conversation, but their input is either ignored until the current generation completes or queued for the next turn. This mismatch between the conversational affordance and the actual capability creates deep frustration — users report 'I told it to stop but it kept going' or 'I corrected it but it ignored me.' The fix: make the streaming phase feel like 'the AI is talking' \(one-way\), not 'we're in a dialogue' \(two-way\). The tradeoff: this makes the interaction feel less conversational and more turn-based, but it sets correct expectations and prevents the frustration of ignored mid-stream input.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T07:22:39.728636+00:00— report_created — created