Harvard Study: Claude Outperforms Human Experts in Peer Review

A study by Harvard University's Gary King and Steven Worthington compared peer review reports written by Claude Opus 4.8 with those from human experts, using 124 papers.

Claude won across all seven review dimensions measured. When GPT-5.5 acted as judge, Claude scored 8.7 on average versus 6.0 for humans, and human reviewers were found to be nearly three times more likely to make factual errors.

The authors said the results held up after ten types of robustness checks.