Posting studies on preprint servers such as bioRxiv is common practice in many scientific fields.Credit: Michael Szebor/Nature
The central conclusions of biomedical preprints rarely change following peer review in a journal1, according to a study posted on the preprint server bioRxiv this month. The research also found that studies that appeared first as preprints are retracted at roughly half the rate of papers that did not appear online before being in a peer-reviewed journal. The authors say the findings suggests that preprints are a reliable source of information, although some scientists say the finding should be interpreted more cautiously.
Posting preprints is common practice in science these days, but Ruslan Rust, a neuroscientist at the University of Southern California in Los Angeles, says that he often hears fellow scientists say that they are unreliable. In his experience, peer review does not typically lead to major changes in the contents of a study. Rust wanted to see if this held true across different fields of biomedical research.
Using a large language model (LLM), Rust and his colleague extracted the main scientific conclusion from the abstracts of 72,644 biomedical manuscripts that were first uploaded to bioRxiv between 2018 and 2025. The model then assessed how much the abstracts had changed compared with their eventual peer-reviewed versions. The study, which has not been peer reviewed, reports that 39.9% of main conclusions were unchanged between the preprint and journal-published abstract, another 50% underwent only minor revisions. Slightly more than 10% went through major changes.
When conclusions did change, they were more likely to become more cautious than more confident after peer review, the study found. Around 8.4% of main findings adopted more cautious language after peer review, whereas 4.2% used more confident wording.
The extent of revision also varied across disciplines. Major changes occurred in only 7.2% of bioinformatics papers compared with 17.5% of microbiology studies.
The authors also found that the frequency of major revisions declined over time, reducing from 17% among papers posted in 2019 to 5.7% in 2024.
Julian Sienkiewicz, who studies artificial intelligence tools and data exploration at the Warsaw University of Technology, says the decrease in major revisions over time could indicate that peer reviewers are overloaded and might not be reading papers thoroughly.
Rust suggests that the decrease reflects a change in how people use preprints. In earlier years after bioRxiv was launched, particularly during the COVID-19 pandemic, scientists were under pressure to post their findings in a short time. This meant many papers had to undergo major revisions before publication, he added. In the past few years, some manuscripts might have already included reviewer feedback in the first preprint version that was posted online, he adds.
Fewer retractions
Rust and his colleague also found that papers that first appeared as preprints were retracted at a rate of 8.1 per 10,000 papers, compared with 18.7 per 10,000 comparable papers that had never been posted as preprints.
The authors caution that the comparison is observational, based on relatively few retractions and does not prove that posting a paper as a preprint reduces the likelihood of a retraction.
Reacting to the findings on the professional-networking platform LinkedIn, some researchers noted that preprints are subject to strong selection bias depending on who posts them and which studies are posted.
