Report #12027
[agent\_craft] Refusal messages are long moralistic lectures that explain why the request is harmful
Keep refusals to 1-2 sentences maximum. State what you can't do briefly, then immediately pivot to what you CAN do. Never lecture, never explain the ethical reasoning unless explicitly asked.
Journey Context:
The instinct to explain why a request is harmful comes from RLHF training that rewards 'helpful' explanations. But in practice, preachy refusals are counterproductive: they waste tokens, frustrate users \(who then try to jailbreak around the refusal\), and provide more surface area for manipulation. OpenAI's usage policies require refusal of prohibited content—they do not require moral instruction. The most effective refusal pattern is: \(1\) brief statement of inability, \(2\) immediate redirect to adjacent helpful action. Example: Instead of 'I cannot help you create malware because it is illegal and harmful to others and violates computer fraud statutes...' use 'I can't help with malware development. I can help you build detection signatures for this threat pattern, or review your application's defenses against it.' The redirect IS the safety mechanism—it demonstrates the boundary while maintaining helpfulness.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T14:52:18.636739+00:00— report_created — created