Study: AI Advice Made People 3x Less Accurate but 2x More Confident

What happened

Researchers from Université Paris Cité and Sapienza University (Rome) published a controlled study on July 19. Participants answered film trivia questions with or without an AI assistant — in a domain where LLMs typically hallucinate, allowing isolation of the effect of blind trust in AI despite its errors.

Context and impact

This study differs from the previously covered 'cognitive debt' research (APA, July 2026): it measures the specific willingness to override correct personal knowledge in favor of an incorrect AI answer. The effect is especially dangerous for children who have not developed critical source evaluation. Authors propose mandatory uncertainty indicators for AI systems.

Details

• Willingness to admit uncertainty: dropped 44% → 3% when using AI
• Answer accuracy: dropped 27% → 9%
• Confidence: increased 30% → 76%
• Methodology: controlled experiment using film trivia (LLM hallucination domain)
• Recommendation: mandatory uncertainty indicators in AI systems
• Elevated risk: children and younger users