I have often wondered why, among three reviews of the same biomedical research manuscript, one reviewer may dismiss nearly everything the authors have done, another may offer a moderate assessment, and a third may strongly favor publication. Human review remains essential, but human judgments can vary widely. Some disagreement is useful because reviewers notice different things. In other cases, the differences may reflect misunderstanding, fatigue, disciplinary preference, or bias. The handling editor must then decide which criticisms are scientifically important and whether the authors should be invited to revise. Such decisions can have significant consequences because published conclusions may ultimately influence clinical practice, public health, toxicological risk assessment, and regulation.
Pressure to publish is also not new. When I was a postdoctoral researcher in the early 1990s, my adviser expected approximately two first-author papers each year. The first paper in a new area usually took longer. As graduate students, however, we sometimes spent at least three years producing the research needed for one full paper. Expectations have changed over the years, while the numbers of researchers, journals, collaborations, and methods for generating and analyzing data have increased. Generative AI arrived in a publication system that was already expanding rapidly.
In his recent Science feature “Ready for the robot reviewers?” Jeffrey Brainard describes the growing burden on editors and reviewers and asks whether artificial intelligence can help. The graph accompanying the article also deserves closer examination.
The graph shows annual scientific publications, rather than manuscript submissions, increasing from slightly more than 1 million in 1990 to nearly 9 million in 2025. The rise is not linear and appears to be accelerating, although the underlying data would be needed to characterize the relationship reliably. What I do not see is an obvious break around 2022, when ChatGPT became widely available: nothing resembling a phase transition, a DNA melting transition, or the marked change in slope sometimes seen in an Arrhenius plot. Instead, the graph shows a long period of accelerating growth that began well before generative AI. AI may have contributed to the most recent increase, but this graph alone cannot tell us how much.
Real physician voices, twice a week
Free, and one click to unsubscribe.
The problem is not simply the number of manuscripts. The same researchers asked to review them are also writing grant proposals, evaluating other grants, teaching, mentoring students, serving on committees, administering research programs, and maintaining their own laboratories. I have previously argued that poor peer review may reflect not simply the shortcomings of individual reviewers but also the broader pressures shaping academic work. One analysis estimated that researchers worldwide devoted more than 100 million hours to peer review in 2020. The estimated value of the time contributed by reviewers in the United States alone exceeded $1.5 billion.
The “2026 Future of Peer Review” report, cited by Brainard, documents growing reviewer fatigue and changes in reviewer behavior. Drawing on that report, Brainard notes that editors needed an average of 4.5 invitations to obtain one completed review in 2025, nearly twice the number required in 2018. Delayed responses and unevenly thorough reviews are therefore not surprising.
For peer review, a hybrid model seems preferable to either human-only or AI-only evaluation. AI could perform an initial technical assessment by flagging possible inconsistencies in calculations, statistics, citations, reporting standards, and the relationships among the text, tables, and figures. The resulting report could then be provided to reviewers as part of the review materials. Providing the same technical assessment to every reviewer would standardize one part of the process and avoid the additional variability created when reviewers independently use different models, prompts, or tools.
Such a system would need safeguards. The AI tool should operate within a secure, journal-controlled environment that protects the confidentiality of unpublished manuscripts and their underlying data. Reviewers should also form an independent initial impression before relying on the automated report; otherwise, the report could anchor every reviewer to the same potentially erroneous concern. Human experts could then verify the AI-generated observations and evaluate novelty, scientific importance, methodological appropriateness, and whether the conclusions are justified. Any AI-generated criticism that could influence the editorial decision would require human verification. Journals should disclose when automated screening has been used, preserve an auditable record of the resulting report, and allow authors to respond to any AI-generated concern that materially affects the editorial decision.
This distinction matters especially in medicine, public health, and toxicology. Published findings may influence clinical decisions, risk assessment, regulation, and public understanding. An automated technical assessment may help improve a manuscript, but it cannot decide how much weight the evidence should carry.
AI will not remove the difficulties of peer review. It may introduce its own biases, overlook discipline-specific considerations, or make reviews produced with similar models too uniform. AI should serve as a second reader: systematic, tireless, and useful, but also fallible. It may direct human attention to possible problems, but it must not become the final judge. Scientific judgment and editorial responsibility must remain with reviewers and editors.
Rao M. Uppu is professor of environmental toxicology and chemistry at Southern University and A&M College in Baton Rouge, Louisiana, where he has served since 2002. He is also an adjunct professor of chemistry and pathobiological sciences at Louisiana State University.
Trained in biochemistry, physical organic chemistry, and free radical chemistry, his research spans bioanalytical methods, biomarker discovery and validation, chemical toxicology, environmental chemistry, computational genomics, reactive intermediates, and the molecular mechanisms of disease. His scholarship is indexed on ORCID, ResearchGate, and Google Scholar.
Dr. Uppu is a Fellow of the American Association for the Advancement of Science (AAAS), the Royal Society of Chemistry (RSC), the Royal Society of Biology (RSB), the Royal Society of Medicine (RSM), the Royal Society for Public Health (RSPH), and the Academy of Toxicological Sciences (ATS).
In addition to his scientific research, he writes reflective essays on scientific culture, mentorship, peer review, ethics, and the human dimensions of academic life. He shares updates on LinkedIn.
