What can AI actually do for peer review right now?
AI is already capable of handling the most time-consuming parts of peer review: screening, summarizing, and even predicting outcomes. In a study using 3,881 radiotherapy patients, AI models screened cases to flag those at risk of treatment interruptions—a task that normally requires careful human review—with accuracy up to 78% for certain patient subgroups [1]. This means reviewers could prioritize their attention on the most complex or high-risk cases, rather than reading every detail of every submission.
Another study trained AI models on thousands of peer-review comments and found that fine-tuned models could predict whether a manuscript would be accepted or rejected with 91% accuracy, based solely on the text of the reviews [3]. That's not just a parlor trick—it suggests AI can learn to recognize the subtle signals that indicate a paper's quality, which could help editors triage submissions more efficiently. However, the same study noted that the models were less accurate when not fine-tuned, so the technology still requires careful setup and validation.
Will AI actually make peer review faster and fairer?
The biggest promise is reducing reviewer workload and speeding up decisions, but the evidence is mixed on whether AI will also make reviews fairer. On the one hand, AI can help non-native English speakers write clearer reviews and can catch inconsistencies in data and methods [2]. On the other hand, AI models can perpetuate existing biases—for example, GPT-4 has been shown to produce racial and gender stereotypes in medical contexts [2]. So while AI might save time, it could also amplify unfairness if not carefully monitored.
A 2021 study that trained an AI on 3,300 papers and their reviews found that AI could predict review scores from text alone, but it also revealed correlations that might indicate bias in the review process itself [5]. This is a double-edged sword: AI can uncover bias, but it can also replicate it. The authors of that study and others caution that AI should be used to assist, not replace, human reviewers [4][5]. In practice, that means AI might flag a paper as 'likely to be rejected' or 'needs more statistical review,' but a human still makes the final call.
What will actually change in the next two years?
The most likely change is that AI tools will become standard in the initial screening stage, not in the final decision. Journals are already experimenting with AI to check for plagiarism, formatting, and even to summarize content [5]. Over the next two years, we'll likely see more journals adopt AI-assisted screening tools that flag potential issues for human reviewers, similar to how the radiation oncology study used AI to flag high-risk cases [1]. This could save millions of working hours, as the 2021 study estimated [5].
But there's a catch: confidentiality and accountability. Many publishers now require disclosure when AI is used, and some prohibit uploading unpublished manuscripts to AI tools due to privacy risks [4]. So while AI might help with the grunt work, the human reviewer will remain responsible for the actual judgment. The consensus across these studies is that AI will be a 'co-pilot' for reviewers, not an autopilot [2][4][5].
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2021 to 2026, 4 from 2024 or later, 3 in Q1 journals, collectively cited 242 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 30 papers retrieved from a database of over 500 million.
Sources used in this answer
Artificial Intelligence-Assisted Peer Review in Radiation Oncology
In a study of 3,881 radiotherapy patients, AI models (LASSO and neural network) screened cases for treatment interruptions with accuracy up to 78% for certain patient subgroups, suggesting AI can aid peer review by flagging high-risk cases.
Artificial-intelligence-based peer reviewing: opportunity or threat?
An editorial in The Lancet Global Health discusses both benefits (e.g., helping non-native speakers, identifying inconsistencies) and risks (e.g., bias, superficial reviews) of AI in peer review, citing that 57.4% of authors found GPT-4 feedback helpful and 82.4% considered it better than some human reviewers.
Application of large language and artificial intelligence modeling in the prediction of peer-review outcomes
Fine-tuned GPT-4mini and GPT-3 models predicted manuscript acceptance with 91% AUC from reviewer comments alone, while untrained models performed much worse (AUC 0.67-0.70), showing the importance of fine-tuning for accurate predictions.
Artificial Intelligence in Scholarly Peer Review: Ethical Considerations, Current Practices, and Future Implications
A 2026 report concludes that AI should be augmentative, not substitutive, in peer review, citing concerns about confidentiality breaches, algorithmic bias, and undisclosed AI use, with major publishers requiring disclosure and human accountability.
AI-assisted peer review
A 2021 study trained an AI on 3,300 papers and their reviews, showing it could predict review scores from text alone and uncover potential biases in the review process, while also highlighting ethical concerns about algorithmic bias.