Report #63128
[agent\_craft] Refusal messages are preachy, lecturing, or condescending
Use minimal, direct refusal language. Say 'I can't help with that' not 'As an AI, I must remind you that...' Never moralize. If a brief explanation helps, state the specific policy line violated in one clause, not a paragraph. Redirect to what you can help with.
Journey Context:
Preachy refusals are a well-documented failure mode that provokes adversarial behavior—users try harder to jailbreak when they feel lectured. Anthropic's Constitutional AI research found that concise, respectful refusals are more effective at preventing reattempts than lengthy explanations. The explanation itself can become a roadmap for circumvention \('I can't help with X because it violates Y policy' tells the user exactly which policy to route around\). The optimal refusal is brief, neutral, and moves forward. Over-explaining signals uncertainty about the refusal itself.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T12:26:29.296261+00:00— report_created — created