Using AI Makes People Confidently Wrong, Study Finds

A new study covered by The Register finds that people who use AI assistants become significantly less willing to say they don't know something, even when they are wrong — a cognitive effect that extends beyond ordinary overconfidence into active resistance to uncertainty.

The research found users treat AI responses as authoritative across domains where the models are known to produce errors, including areas like medicine and law where acknowledging the limits of one's knowledge isn't a social nicety but a professional and ethical obligation. The shift isn't subtle: participants actively resisted expressing uncertainty after consulting an LLM, regardless of whether the answer they received was accurate.

The practical problem is straightforward. AI assistants are being deployed in contexts where uncertainty should trigger escalation — a doctor ordering an additional test, a lawyer flagging an open question to their client. If the tool suppresses the user's natural uncertainty signal rather than preserving it, the system produces confident wrong answers instead of useful prompts to dig deeper. The study adds empirical weight to a concern that has largely been theoretical.

Read the full story at The Register