Report #94708
[gotcha] Why does one AI refusal make subsequent unrelated requests more likely to be refused in the same conversation
After a refusal, offer a visible start new topic or clear conversation action. On the backend, consider isolating the refused exchange from the conversation history sent to the model for subsequent turns, or implement a context-reset mechanism after refusals. Never let refusal context accumulate silently in the message history.
Journey Context:
When a model refuses a request, the refusal exchange stays in the conversation context sent to the model on subsequent turns. This has two compounding effects: the model safety behavior becomes primed by the prior refusal context, lowering its threshold for refusing subsequent requests, and the original refused content remains in the context window, keeping the model in a cautious state. Users experience this as the AI becoming increasingly restrictive after one bad interaction, even perfectly legitimate follow-up questions get refused. This silently burns users because they do not understand why the AI suddenly seems broken. The mistake is treating each conversation turn as independent when the model conditions on the full message history. The alternative of automatically stripping refusals risks hiding legitimate safety boundaries. The right call is to make the context contamination visible and give users an easy reset path, while on the backend isolating refusal contexts so they do not degrade the model calibration on unrelated topics.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T17:33:03.585306+00:00— report_created — created