Some journals prohibit AI-written reviews; geneticist Anshul Kundaje refuses to review for them, while the head of the aging research funding program Impetus Grants considers AI more capable than any individual human reviewer for many applications

Some journals prohibit AI-written reviews; geneticist Anshul Kundaje refuses to review for them, while the head of the aging research funding program Impetus Grants considers AI more capable than any individual human reviewer for many applications

On October 27, Heng Li of Harvard Medical School endorsed the bans. On October 28, Kundaje and Ruslan Rust, whose group tested AI reviewers, challenged his position. On October 29, Martin Borch Jensen described the experience at Impetus Grants.

Heng Li, who developed BWA, a program for analyzing DNA sequencing data, writes in a blog post that AI likely reviewed two manuscripts from his group. Li considers a review AI-generated if the reviewer starts with AI before reading the paper carefully, even if they later edit the text: AI has already shaped their judgment. He accepts AI assistance if a human writes the first draft.

Li argues that “AI often misses the big picture because it lacks taste, despite having the knowledge.” He submitted three published papers on the same topic to an AI review website. The highest score went to a paper whose main conclusion he considers half wrong; the lowest went to the paper he prefers.

The publisher Springer Nature prohibits reviews written entirely by AI, while Science prohibits submitting a manuscript to AI to generate or supplement a review. Since June 2023, the US National Institutes of Health (NIH) has prohibited grant reviewers from using AI to analyze applications or write reviews because these tools provide no guarantees about where the data goes.

Anshul Kundaje, an associate professor of genetics at Stanford University, disagrees with Li: on average, an AI review is better than a review by the average human reviewer, particularly in specialized and interdisciplinary fields. Kundaje recommends a sequence that Li accepts but Science prohibits: first read the paper and write a review, then use AI to supplement it. Kundaje states: “I will not review for journals that prohibit the use of AI.” He plans to review preprints (papers that have yet to undergo journal peer review) publicly, with AI assistance, calling the journal system “completely broken and irrelevant.”

Rust cites figures from his group's preprint, which repeated a 2008 experiment using AI models. In the original experiment, 607 reviewers received three clinical trial reports containing deliberately inserted errors. Reviewers who had received no training detected about 27% of the major errors. When 15 models read the same reports, they detected 53–96% of the major errors. The reports have been publicly available since 2004, so the models could have memorized them. The authors therefore inserted the same types of errors into reports of 58 trials from 2026: the models detected 58–99% of the major errors in these reports. In response to concerns about confidentiality, Rust argues that preprints contain no secrets and that unpublished work can be reviewed using open models running on one's own computer. Four of the five authors, including Rust, are affiliated with ResearchHub, a platform implementing AI-assisted peer review.

Variation between models also appears in another group's preprint: four models reviewed 200 published papers from 2026. Of the 800 reviews, 75,6% recommended major revisions or rejection; for individual models, that proportion ranged from 17,5% to 100%.

Borch Jensen, who heads Impetus Grants, writes that the program has tried both human and AI reviewers. For many applications, he considers the models more capable than any individual human: they can identify problems across statistics, biology, and epidemiology at once. Humans have a better grasp of tacit knowledge: they know about “all the unpublished experiments, researchers' motivations, ideas that look good on paper…”.