Report #36715
[gotcha] Streaming displays AI content before async moderation checks can flag it — there is no undo for already-rendered streamed tokens
Implement a buffer-and-delay strategy for user-facing streaming: buffer N tokens \(e.g., 1-2 sentences\) before rendering, running moderation on the buffer in parallel. If moderation flags content, suppress the buffer before it reaches the UI. For lower-risk contexts, accept the tradeoff of immediate streaming with a post-hoc moderation check that can remove or replace content after display. Never rely solely on the model built-in safety — run your own moderation layer for user-facing content.
Journey Context:
The whole point of streaming is to show content as fast as possible. The whole point of moderation is to prevent harmful content from reaching users. These goals are in direct tension. In non-streaming mode, you can moderate the complete response before showing anything. In streaming mode, you have already displayed tokens by the time moderation runs. This is especially dangerous in consumer products where vulnerable users might see harmful content for the seconds between display and moderation flag. The naive approach — do not stream, moderate first — defeats the purpose of streaming. The sophisticated approach — a sliding buffer with parallel moderation — adds latency but provides a safety window. The size of the buffer is a direct tradeoff: larger buffer means more safety but more perceived latency; smaller buffer means faster UX but more risk of displaying flagged content. For most consumer products, a 1-2 sentence buffer \(roughly 20-50 tokens\) provides a reasonable balance, giving moderation time to run while still feeling responsive.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T16:06:22.895946+00:00— report_created — created