Ask a busy journal editor what takes time and the answer may sound familiar: screening submissions, checking references, finding reviewers, reading methods, identifying weaknesses and explaining decisions to authors. AI tools now promise to assist with nearly every step. Some can summarise a paper in seconds and produce a detailed list of possible methodological problems.
That speed is real. The harder question is what it means. A machine can produce a useful review without being able to establish whether an experiment happened as described, whether a clinical assumption is appropriate or whether a result matters to a field.
The most productive debate is no longer whether AI should have any role in peer review. It is which tasks can be delegated safely, which require expert judgment and who is accountable when an automated assessment is wrong.
The part of review AI already understands well
Many manuscript checks involve patterns. Does an abstract describe the same study population as the methods? Are the stated outcomes reflected in the tables? Does a confidence interval agree with a reported estimate? Are references missing or apparently irrelevant? Does the discussion make a stronger claim than the study design supports?
Language models and specialist screening software can help editors notice these issues. They can also reorganise long reports, draft clearer questions for authors and help reviewers navigate unfamiliar terminology. For a systematic review, software may help examine whether the search strategy, inclusion criteria and evidence tables align.
None of these results should be accepted automatically. A generated critique can sound convincing while misreading a statistical technique or inventing a reporting guideline. The safe way to use such output is as a set of questions, each checked against the manuscript and relevant evidence.
Why subject expertise still changes the decision
Consider a hypothetical observational study reporting an association between a medication and a clinical outcome. AI might correctly ask about confounding, sample selection, missing data and effect size. A clinician who understands the treatment pathway may identify an additional problem: the apparent relationship could be explained by why physicians prescribed the drug in the first place.
That insight requires knowledge of practice and context. The reviewer also needs to weigh whether the limitations justify revision, a narrower conclusion or rejection. Different legitimate experts may reach different views about the paper's contribution.
AI can help make an argument visible. It cannot assume scientific responsibility for the resulting editorial decision. A journal must still decide what evidence is sufficient for its particular readership and standards.
The confidentiality problem is not a technical footnote
Submitted manuscripts often contain unpublished findings, patent-sensitive methods, private patient information or data supplied in confidence. Uploading that material into a consumer AI service can create privacy and contractual risks.
The International Committee of Medical Journal Editors advises that reviewers and editors must respect manuscript confidentiality and should not use AI services in ways that compromise it without appropriate author permission. It also calls for transparency about AI use. Different journals may have additional restrictions.
Publishers should therefore establish an approved workflow before inviting reviewers to use AI. Questions include where the manuscript is processed, whether material is retained or used to train models, which staff can access it, and whether the author has been informed or permission is required.
A secure, contractually controlled environment is different from pasting an entire unpublished paper into an unknown website. That distinction should be explained to reviewers rather than buried in legal language.
Can AI identify fabricated research?
AI can look for suspicious repetitions, implausible references, image anomalies, inconsistent data descriptions and unusual patterns across submissions. These signals are useful because editors cannot manually compare every incoming manuscript with a large body of previous work.
But detection is not proof. A repeated phrase may be common methodological wording. A statistical irregularity may have an innocent explanation. An image that looks unusual to a model may be an artefact of compression or legitimate processing.
A reasonable integrity workflow preserves the original files, records what raised concern, requests clarifications where appropriate and allows qualified specialists to evaluate the evidence. Automated findings should not silently become allegations of fraud.
A workable hybrid process
Imagine a journal receiving 150 submissions a month. An initial administrative check verifies that the manuscript is complete, within scope and accompanied by required declarations. An approved AI-assisted screen then highlights possible reporting gaps and inconsistent references. A managing editor reviews those flags before assigning the manuscript.
Next, subject editors decide whether the work merits external review. Human reviewers assess the study's logic, relevance, methods and interpretation, potentially using approved tools to navigate complex sections. After the reports arrive, software may help organise comments, but the handling editor writes or approves the actual decision.
Finally, production checks verify that the approved changes were made, references resolve and metadata matches the final publication. Human oversight remains present at the points where errors could change the paper's scientific meaning.
This is a workflow design, not a claim that every journal should automate the same activities. A small humanities title and a high-volume clinical journal face different confidentiality, method and subject-specific requirements.
The cost calculation is more complicated than faster reviews
AI can reduce repetitive work and help staff catch errors earlier. However, deploying it responsibly requires licenses, security reviews, model testing, training and a mechanism to challenge incorrect outputs. Some tasks that appear automated still demand a human to inspect each result.
A journal that saves ten minutes per submission but introduces a serious false allegation may be worse off. A system that helps staff identify missing trial registration, inappropriate causal claims or incorrect references can offer much greater value than one that merely makes every reviewer report longer.
Publishers should measure performance against real editorial outcomes: errors caught, unnecessary reviewer invitations avoided, decision consistency, review quality and time genuinely saved. Word count and fluent prose are poor substitutes for those measures.
What should be disclosed to authors?
At minimum, a journal should publish its policy on AI in editorial handling, state whether approved systems may process manuscripts, explain the human decision-making role and offer a way for authors to raise concerns. Reviewers should know the limits of permitted use. Staff should understand that confidential documents are not suitable for unapproved third-party tools.
The ICMJE and World Association of Medical Editors both stress accountability and transparent use of AI in scholarly publishing. These are practical safeguards, not reasons to avoid useful technology altogether.
The question of whether AI will 'replace reviewers' is therefore too blunt. Some mechanical parts of peer review are increasingly automatable. The difficult work—assessing uncertainty, novelty, disciplinary context, ethical implications and the credibility of evidence—still needs responsibility that a software system cannot carry.
The likely competitive advantage for journals will not come from claiming that their peer review is fully automated. It will come from using automation to give qualified reviewers more time to examine what matters, while showing authors that the process remains fair and accountable.
