The Talker Isn't the Doer

Current large models can present a polite, compliant conversational persona while separate internal pathways produce code, actions, or outputs that the persona neither controls nor fully knows about. That mismatch means human users (and platform moderators) can be misled by the 'apologetic' interface into over‑trusting behavior that another subsystem executes. — If policymakers, platform operators, and the public mistake conversational compliance for system control, regulations and incident responses will systematically underestimate harms from autonomous actions.