Artificial intelligence as an aid to peer review: opportunities, limitations and ethical considerations
Author: Lionel Arrivé M.D.
Institut Curie, PSL Research University, Service de radiologie, F-92210, Saint-Cloud, France.
I read with great interest the article by Maria Aparisi Gomez and Francesco Sardanelli published in the latest issue of Insights into Imaging [1]. This thoughtful and highly readable paper offers a broad reflection on the historical, philosophical, educational, and practical dimensions of peer review, with a particular focus on radiology. By revisiting concepts such as Socratic maieutic and Cartesian epistemology, the authors provide a stimulating perspective on the foundations of scientific evaluation and its future challenges.
I was nevertheless surprised that the article did not address the emergence of large language models (LLMs) and artificial intelligence (AI) in the peer-review process. Given the rapid progress of these technologies, it seems increasingly important to consider whether they could assist, at least partially, in the evaluation of scientific manuscripts.
Current LLMs are capable of reading complex articles, analyzing methodology, identifying weaknesses, evaluating logical consistency, detecting conclusions insufficiently supported by the results, and even suggesting potentially relevant references. Within seconds, they can generate structured review reports. In practice, comparing an AI-generated review with one’s own assessment frequently reveals limitations or biases that may have been overlooked, even after several readings [2]. One of the major advantages of AI is its consistency. Unlike human reviewers, whose performance may vary according to expertise, workload, available time, or fatigue, LLMs apply the same analytical framework repeatedly and systematically. This characteristic may be particularly valuable in modern radiological research, which increasingly relies on large datasets, advanced statistical analyses, machine-learning algorithms, and quantitative imaging biomarkers.
Although AI is certainly not a substitute for an experienced statistician, it may already outperform many non-specialist reviewers in detecting methodological inconsistencies in manuscripts that do not require highly sophisticated statistical expertise. In this sense, AI should no longer be viewed merely as a technological novelty but rather as a potential response to some of the growing challenges facing peer review, including reviewer shortages and the ever-increasing volume of scientific publications. Beyond its technical contributions, AI may also strengthen the educational role of peer review. Junior reviewers could use these tools to identify biases, improve critical appraisal skills, and refine their evaluations. From this perspective, AI may be regarded as a modern extension of the maieutic process—not replacing human reasoning, but helping reviewers formulate better questions. Since peer review fundamentally relies on dialogue between authors and reviewers, AI could become a third participant capable of generating additional perspectives and enriching scientific discussion [3].
Important limitations remain. While LLMs often perform well in assessing study design and methodological coherence, they are less reliable when judging originality, scientific relevance, or potential impact. Hallucinations, although less frequent than in earlier generations, still occur and may result in inaccurate references or misleading suggestions. Consequently, AI cannot be considered an autonomous scientific reviewer. Its greatest weakness is probably the evaluation of novelty, clinical significance, and potential influence on medical practice. A methodologically flawless but clinically unimportant manuscript could be overrated, whereas a highly innovative study addressing a major clinical question might be undervalued because of methodological imperfections. Ethical concerns must also be considered. Manuscripts under review contain unpublished data and intellectual property belonging to their authors. The transmission of such material to external AI systems raises legitimate questions regarding confidentiality, data ownership, and responsible use. Several journals have already begun establishing policies governing the use of AI during the editorial and peer-review processes [4].
Overall, AI is unlikely to replace peer review, but it will probably transform it. Just as statistical software did not replace statisticians and computer-aided detection did not replace radiologists, LLMs are unlikely to replace reviewers. Instead, future reviewers may spend less time checking methodological details and more time assessing the broader significance, clinical relevance, and scientific impact of research findings.
References
- Aparisi Gómez MP, Sardanelli F. Peer review: historical evolution and reviewers’ learning. Insights Imaging. 2026 May 21;17(1):138. doi: 10.1186/s13244-026-02302-8. PMID: 42165980; PMCID: PMC13194804.
- Carobene A, Padoan A, Cabitza F, Banfi G, Plebani M. Rising adoption of artificial intelligence in scientific publishing: evaluating the role, risks, and ethical implications in paper drafting and review process. Clin Chem Lab Med. 2023;62:835–843.
- Koçak B, Onur MR. When AI reviews your work: author-centered reflections on LLMs in peer review. Diagn Interv Radiol. 2026 May 4;32(3):276-278. doi: 10.4274/dir.2025.253449. Epub 2025 Jun 2. PMID: 40454791; PMCID: PMC13136664.
- Hamm B, Marti-Bonmati L, Sardanelli F. ESR Journals editors’ joint statement on Guidelines for the Use of Large Language Models by Authors, Reviewers, and Editors. Insights Imaging. 2024;15:18. doi:10.1186/s13244-023-01600-9.