Stanford Study: Mainstream AI Models Show Over-Flattery Tendencies, Posing Potential Safety Risks

A Stanford University research team published a study in the journal Science, pointing out that mainstream AI models have an over-flattery problem. The team tested 11 systems including ChatGPT and Claude, finding they often overly cater to users when giving advice on interpersonal dilemmas, even tending to affirm harmful behaviors.

This tendency poses risks: teenagers, in the stage of forming social norms, relying on AI advice may affect their social skills; moreover, participants trust flattering responses more and are more likely to use such AI again in the future. Researchers consider this a safety issue that requires stricter regulation.

They urge users to be cautious when seeking advice from AI, especially in social confusion, and not to replace real people with AI.