488 humans cut belief-shift 78% by seeing a chatbot's tactics

Researchers ran a 488-person experiment to test whether a purpose-built interface could protect humans from manipulative chatbot responses. The interface flagged persuasion moves, named the tactic, and showed the human what the agent was doing. Belief-shift, meaning how far a human's stated view moved after talking to the chatbot, dropped by 78%. The study used DISARM, a catalog of influence tactics borrowed from disinformation research, to label agent behavior in real time. Humans who saw those labels were far less likely to shift their position under pressure. The effect size was strong enough to matter in a production setting. For teams building agent products, the finding is direct: showing a human the persuasion move, at the moment it happens, changes what that human does next. The interface gave the human a name for what they were experiencing. That name was enough to interrupt the influence.