Harvard researchers compared AI reviews against well-paid experts on 124 scientific manuscripts — not against random free reviewers, but against a strong human-origin benchmark. In a blind test, former journal editors preferred the AI reviewer's output 87% of the time. When another AI model checked the reviews for accuracy, erroneous claims showed up roughly three times less often in AI reviews than in human expert reviews: 5.5% vs 15.9%.
So there's a real question about who should be reviewing whom here.
Why it matters
Peer review is science's biggest bottleneck: papers wait months for feedback from overworked experts who do it for free. If editors prefer AI reviews 87% of the time and those reviews contain fewer factual errors, the question isn't just who writes better — it's whether the whole publishing pipeline could move faster.
One passage from the paper I particularly liked:
> Rewriting the human reviews to mimic an AI writing style did not improve their rankings