Harvard researchers find Claude-written paper reviews beat paid human reviewers, judged by AI and by journal editors
Gary King and Steven Worthington compared 134 human reviews from their rapidPeer service with Claude Opus reviews; AI won every pair under a GPT-5.5 judge and 87% when three former journal editors judged.
Why it matters
The result comes from the people who run rapidPeer, a paid double-blind review service at Harvard's Institute for Quantitative Social Science, and King says the service will now have to change what it does. The paper has not itself been peer reviewed, and the human editors judged only 10 randomly chosen pairs. The authors also warn that AI reviews could push science toward conventional, majority views. Watch whether journals formally test AI review and whether the finding holds in independent replication.
Key facts
- 134 human reviews of 124 papers, 2019 to 2026
- AI reviews written with Anthropic's Claude Opus
- GPT-5.5 judge preferred AI in 100% of pairs
- Three former journal editors preferred AI in 87%
- Editors judged 10 random pairs
- Paper not yet peer reviewed