More Confident Human Review Can Remain Wrong
Recchia and colleagues’ extended preprint v1 (2025-05-17) tested difficult two-answer questions in law, medicine and the Lojban language. Study 1 detected no overall accuracy benefit from initial opposing AI-generated arguments before consultation, nor evidence that assisted humans outperformed the model alone. Exploratory comparisons found higher accuracy with arguments when the model was wrong. Study 2’s preregistered analyses found no significant overall differences with versus without AI-generated research guidelines alongside one AI advisor. After roughly 20 minutes’ web research, both groups increased confidence in that advisor’s answer on average, including when wrong. Confidence and correct checking differ. Simple protocols, often unhelpful guidelines and Study 2’s changed instructions/incentives limit generalization; these results neither establish universal oversight failure nor test new-discovery review.
0
1
Tags
Scientific Oversight of AI-Assisted Research
1Cademy