# The Ethics of Using AI Tools in Peer Review: What's Allowed In most cases, uploading an unpublished manuscript to a generative AI tool as part of your peer review is not allowed, because it breaches the confidentiality you agreed to when you accepted the review. Light assistance such as grammar and spelling correction in your own written report is widely accepted, and some journals run their own approved AI tools, but using a chatbot to evaluate the science itself is prohibited by the major editorial bodies. If you have been invited to review a paper and are wondering whether ChatGPT, Claude, Gemini, or a similar tool can help, the answer depends on three things: the journal's stated policy, the confidentiality obligations of peer review, and the ethical principle that a reviewer, not a machine, must take responsibility for the scientific judgment. This article walks through those three factors, explains the reasoning behind them, and gives you a practical checklist you can use before you touch any tool. ## What Does AI in Peer Review Ethics Mean The phrase "ai in peer review ethics" describes the set of rules and moral principles that govern how artificial intelligence tools may be used during the evaluation of manuscripts for publication. It sits inside the broader field of [research ethics](/blog/guides/research-ethics-fundamentals-a-practical-guide-for-early-career-scientists), which covers honesty, confidentiality, accountability, and fairness in the conduct of science, and it is an emerging component of the responsible conduct of research, the formal training framework many institutions require of researchers. A plain-language interpretation looks like this. Peer review is a promise-based system. When you agree to review a manuscript, you promise the editor three things: you will keep the paper confidential, you will judge it fairly and without conflicts of interest, and you will stand behind the accuracy and integrity of your report. The ethics question is simply whether using a particular AI tool helps you keep those promises or breaks them. A short definition you can carry away: AI in peer review ethics is the analysis of when machine assistance with manuscript evaluation is compatible with confidentiality, accountability, and scientific judgment, and when it is not. ### Why the stakes are real Peer review failure is not an abstract concern. Reviews of professional failure in biomedical fields repeatedly identify inadequate peer review as a recurring vulnerability in cases that caused real harm, and identify mandatory, competent peer review as a core safeguard for public trust [1]. In forensic medicine, where peer review protects legal outcomes, regulatory and judicial case analyses list insufficient or superficial peer review among the documented causes of wrongful conclusions [1]. Scientific publishing has the same dependency. A review that misses a fatal statistical flaw, or that is quietly generated by a model, damages the evidence base that clinicians and future researchers rely on. Peer review also has real, substantive power over the final paper. In one published cross-sectional study, the authors state plainly that their statistical model was re-specified during peer review, meaning reviewers shaped the final analytical structure of the article [2]. That is what review is for. A model cannot be held responsible for that kind of editorial influence, and a reviewer who outsources it has neither given it nor received it. ## The Core Rules: What Is Allowed and What Is Not Editorial consensus has crystallized around a set of boundaries. The Committee on Publication Ethics (COPE), the body most publishers follow for publication ethics guidance, states in its position on authorship and AI tools that generative AI should not be used by reviewers or editors to assist in the evaluation or decision-making on manuscripts, and that reviewers should not upload submitted manuscripts to AI tools. The International Committee of Medical Journal Editors (ICMJE) writes similar language into its Recommendations, emphasizing that a reviewer's judgment cannot be delegated to a tool. Individual publishers, including major scientific publishing houses and flagship journals, adopt policies that mirror these positions. The table below summarizes the practical result. | Use of an AI tool by a reviewer | General status | Reason | |--|--|--| | Uploading the full manuscript, figures, or supplementary files to a public chatbot | Not allowed | Breaches manuscript confidentiality | | Using AI to generate the substance of the review or the accept/reject recommendation | Not allowed | Delegates accountable scientific judgment | | Asking AI to check references it "remembers" from training data | Not allowed | High fabrication risk, unverifiable output | | Grammar, spelling, and phrasing cleanup of a review the reviewer wrote | Usually allowed | No confidential content leaves the reviewer's control of substance, and the reviewer authored the judgment | | Using a journal-provided or publisher-sanctioned AI screening tool | Allowed when the journal offers it | Covered by a data agreement with the vendor | | Searching for relevant literature with an AI-enabled discovery tool, without uploading the manuscript | Often allowed | The manuscript itself is not transmitted | | Disclosing any AI assistance to the handling editor | Required in most policies | Transparency preserves accountability | Always check the specific journal's reviewer guidelines, because policies differ in detail even where they agree in spirit. ### What is generally allowed The clearest permitted category is mechanical language assistance applied to text you have already written yourself. If you drafted a review, read the paper carefully, formed your own assessment, and then run your report through a grammar checker to catch typos, that is widely accepted. This practice predates generative AI. Reviewers have used spelling and grammar software for decades, and no major policy treats it as a breach because the reviewer's own analysis is not being replaced and the manuscript is not being transmitted to a third party in any meaningful sense. Some journals and publishers have moved a step further and built AI tools into their own submission and screening systems. These tools operate under data processing agreements with the vendor, which is the ethical difference between the journal's sanctioned workflow and a reviewer pasting text into a public chatbot. Where such a tool exists, the journal tells you what it does. If the journal has not told you a tool exists, assume it does not. Literature discovery is another gray area that often resolves to allowed. If you need to check whether a cited paper exists, whether a claimed method is standard, or whether a prior study contradicts the one you are reviewing, using a search tool with AI features is usually fine, provided you are entering search queries and keywords rather than the manuscript itself. You remain responsible for actually reading whatever sources you rely on. ### What is generally prohibited Three prohibitions appear in nearly every policy. First, confidentiality. A submitted manuscript is the authors' unpublished intellectual property, shared with you under an obligation of secrecy. Uploading it, or substantial parts of it, to a public generative AI service transmits that content to an external company whose data practices you do not control. Most reviewers would never email the manuscript to a stranger, and in ethical terms a public AI tool is a stranger. Some people assume that removing author names solves the problem. It does not, because the obligation covers the whole unpublished work, not just identifying details, and because removed names can sometimes be inferred from the content itself. Second, delegated judgment. Evaluating whether the science is sound is the task you personally accepted. A generative model can produce a fluent critique that looks like a review, but fluency is not evaluation. The model does not reproduce the experiment, check the reagents, re-run the analysis, or examine the raw data. It produces plausible-sounding text based on patterns in its training data. Third, fabrication risk. Generative language models are known to invent plausible but false content, including nonexistent references, wrong statistical claims, and misdescribed methods. In disease classification research, a systematic review of explainable AI methods identifies lack of transparency, trust, and systematic evaluation as the major barriers to deploying black-box models, precisely because outputs cannot always be traced to a verifiable basis [3]. A reviewer who copies an AI-generated claim into a report without verification may be inserting fabricated criticism into the scientific record, or, worse, may be reassuring the editor about a flaw the tool failed to see. ### The gray zone and how to handle it Real cases sit between the bright lines. Examples include using AI to help you structure your notes before writing the review, asking a tool to clarify a statistical concept the paper uses, or having AI reformat your comments into bullet points. Where the manuscript itself is not uploaded and the scientific judgment remains yours, many journals tolerate these uses. The honest procedure is simple: describe what you did to the editor in your confidential comments, and if the journal's guidelines forbid even that, follow the guidelines. Transparency costs nothing. Concealment is what converts a tolerated convenience into an ethical breach. ## Why Confidentiality Is the Center of the Problem Many people focus on accuracy and miss the confidentiality problem, so it deserves its own explanation. When you accept a review, you enter a confidential relationship. Editors enforce it with rules about not sharing the manuscript, not discussing it with colleagues without permission, and not using its content for your own work. These rules exist because unpublished findings are fragile. Premature disclosure can scoop the authors, compromise patent claims, or expose human subjects data. Generative AI services change the nature of data transfer. A chatbot prompt is a transmission to a remote server, and in most consumer services the provider may retain, log, or use submitted text to improve its models unless the account has specific contractual protections. That makes the upload functionally similar to distributing the manuscript. The fact that no human reads the output does not matter ethically, because confidentiality protects the information, not the audience. This is also why the distinction in tool types matters. Precision matters when discussing AI, in peer review and everywhere else. A study of robotic-assisted surgery notes that conflating distinct technologies under one label obscures what the evidence actually supports [4], and the same discipline applies here. A locally installed spell checker on your own machine is not the same class of tool as a cloud-hosted chatbot. A journal's contracted screening system is not the same as a public AI service. When you assess a tool, ask where the text goes, who can see it, and whether the provider is bound by an agreement with the journal or the institution. ## Why Accountability Cannot Be Outsourced The second pillar of the ethics is accountability, and it is the one reviewers most often underestimate. When your report goes to the editor, the editor treats it as your expert judgment. Decisions about publication, corrections, and retractions rest on it. If a model wrote the core of your critique and it was wrong, the harm flows through your name and the journal's decision, and neither you nor the journal can meaningfully audit why the model said what it said. This connects to a broader theme in the medical AI literature. Reviews of explainable AI consistently find that clinicians and evaluators are unwilling to act on black-box outputs they cannot inspect or verify [3]. Peer review carries that logic to its natural conclusion. A review report is itself an audit document. It must be traceable to a person who can defend every claim in it. There is also a competence problem. Strong review requires field-specific expertise, familiarity with the methods, and often the ability to spot what is missing, such as a missing control, an inappropriate normalization, or an overclaimed causal inference. Language models summarize common patterns in published text. Subtle novel flaws are, by definition, not common patterns. A reviewer who relies on the model for substance tends to get a generic critique, which is precisely the kind of superficial review identified as a failure mode in professional contexts where peer review protects consequential decisions [1]. None of this means models are useless to a careful reviewer. Used after the fact, as a readability pass on your own prose, they can genuinely improve reports. The ethical dividing line is not whether electrons from a model touched your workflow. The line is whether the judgment is yours and whether the manuscript stayed confidential. ## AI in Peer Review Ethics Explained Through the Research Ethics Lens It helps to place these rules inside the standard framework of the responsible conduct of research, which organizes professional obligations into recognizable duties. Confidentiality is the duty not to disclose information entrusted to you. Uploading a manuscript to a public tool risks violating it, which is why the prohibitions start there. Integrity and accountability are the duties to report your own judgment honestly and to answer for it. A review signed by you is a claim that the assessment is yours. In one published study, the authors explicitly note that peer reviewers caused their statistical model to be re-specified before publication [2]. Reviews change papers. Whoever makes those changes must be identifiable and answerable, and only a human can be. Honesty about methods is the duty to disclose how work was produced. Researchers increasingly document AI assistance in their workflows. A recent medical education scoping review openly states that its data extraction was performed with an AI-assisted workflow [5]. That transparency is the model for how AI use should be handled in peer review: named, bounded, and visible to the person who depends on it. Fairness is the duty to evaluate work without bias or conflict of interest. Language models inherit biases from their training data, including uneven representation of journals, languages, regions, and methodological traditions. A critique generated by a model may systematically favor mainstream approaches and penalize work outside them, in ways neither the reviewer nor the editor can detect. Human reviewers are also biased, of course, which is why declared conflicts of interest are mandatory. The difference is that a human bias can at least be declared and weighed. ### The author side, for comparison The rules differ depending on which side of the review you sit, and reviewers are usually authors too. Authors, under COPE guidance and the ICMJE recommendations, may use AI tools to help prepare a manuscript in many journals, but must disclose that use in the paper, and AI tools cannot be listed as an author because authorship requires responsibility and accountability that a tool cannot hold. Reviewers, by contrast, face stricter rules: no uploading of the manuscript, no delegated evaluation, and disclosure of any assistance to the editor rather than in the published paper. Editors face their own version: they may use sanctioned AI tools for administrative screening, but publication decisions must rest with accountable human editors. The asymmetry has a logic. Authors own their unpublished work and can choose how to produce it, subject to disclosure. Reviewers hold someone else's property in trust. Holding property in trust imposes stricter limits. ## Common Myths and Related Questions **"Everyone uses it, so it must be fine."** Prevalence is not permission. Policies from COPE, the ICMJE, and major publishers were written because the practice became common, not despite it. **"It is just a spell-checker."** The comparison fails for generative chatbots. A spell checker suggests spelling corrections. A chatbot generates substantive content about the paper. The correct comparison for a chatbot is a colleague who reads the manuscript, and you may not share the manuscript with a colleague without the editor's permission. The discipline of keeping distinct tool types analytically separate, argued explicitly in a surgical technology review that separates AI prediction from robotic execution [4], applies directly here. **"If the AI catches a real error, the review is better."** Occasionally a model will flag something valid. That does not make the workflow acceptable, for the same reason a burglar who finds a gas leak has not performed a home inspection. The process, not the outcome, defines the breach. If you find an error through impermissible means, you cannot honestly report how you found it, which creates a second integrity problem on top of the first. **"I removed the names, so it is anonymous and safe."** Anonymity is not the confidentiality test. Unpublished data and unpublished ideas are what the obligation protects. Anonymized manuscripts still contain the findings. **"AI review tools will catch plagiarism and image manipulation, so they are the solution."** Specialized screening tools, run by journals under contract, do assist with image forensics and text similarity. Those are the journal's sanctioned tools, not your chatbot, and even sanctioned tools flag content for human verification rather than replace it. ## A Practical Checklist Before You Use Any AI Tool in a Review 1. Read the journal's reviewer guidelines and look for an AI or confidentiality section. If a policy exists, it governs. 2. Ask where the text will go. Your own machine, a journal system, or an external cloud service. 3. Ask whether the tool is sanctioned. Journal-provided tools operate under agreements. Public chatbots do not. 4. Keep the manuscript out of the prompt. No full text, no figures, no methods sections, no anonymized versions. 5. Write your own scientific assessment first. If the tool only cleans language in text you authored, you are usually on safe ground. 6. Disclose whatever assistance you used in your confidential comments to the editor. 7. Verify every factual claim in your final report against sources you have personally read. 8. Decline the review if you cannot review it competently within the rules. A prompt decline protects everyone. ## Common Mistakes and Limitations Reviewers make predictable errors in this area, and each has a direct fix. Pasting methods sections into a chatbot to "understand the paper faster." This transmits unpublished work to an external service and is the single most common policy breach. Fix: ask the editor for an extension instead. Trusting AI-generated reference checks. Models fabricate citations and misstate findings. A systematic review of explainable AI in disease classification documents why unverifiable black-box outputs fail to earn trust in professional settings [3]. Fix: verify every claim by reading the source yourself. Copying AI text into the review report. Editors can often recognize generic machine-generated critique, and it fails the accountability test. Fix: every sentence in your report should be one you can defend in a conversation with the editor. Confusing tool categories. Locally run grammar software, journal screening systems, and public chatbots are different classes of tool with different rules. Fix: classify the tool before you use it. Assuming silence means permission. A journal without a written AI policy still has confidentiality rules that predate AI, and those rules already prohibit disclosure to third parties. Fix: apply the underlying confidentiality principle, not the absence of a specific clause. One broader limitation deserves a plain statement. Policies are moving quickly, and this article reflects widely adopted editorial guidance rather than a guarantee for every journal. Individual situations, especially borderline uses or institutional requirements, should be confirmed with the handling editor or your institution's research integrity office before you act. ## What Remains Uncertain Several questions remain genuinely open. It is not settled how journals will verify compliance, since AI-generated review reports are difficult to detect reliably and detection tools carry their own false-positive problems. It is not settled how confidentiality will be handled if publishers negotiate enterprise agreements with model providers, which could change what is permissible in house. It is not settled whether reviewers will be permitted to use AI on published supplementary materials differently from the core manuscript, or how disclosure of minor AI assistance will be recorded in the publication record. What is settled is the principle the major editorial bodies have converged on. The scientific judgment in a peer review must be a human judgment, the manuscript must stay confidential, and any AI assistance that touches the process must be disclosed. Those three commitments are the stable core of ai in peer review ethics, whatever the tools look like next year. ## Frequently Asked Questions ### Can I use ChatGPT to summarize a manuscript I am reviewing? No. Uploading the manuscript to a public chatbot breaches the confidentiality you agreed to as a reviewer, even if you delete the conversation afterward. ### Is it allowed to use AI for grammar and spelling in my review report? In most cases yes, because you apply it to text you already wrote and the manuscript itself is not transmitted. Check the journal's guidelines and disclose the assistance to the editor. ### Do I have to tell the editor that I used an AI tool? Yes. Nearly all current policies require disclosure of any AI assistance in your confidential comments to the editor, even for language help. ### Can AI tools detect plagiarism or image manipulation in a paper I am reviewing? Sanctioned screening tools run by journals can flag text similarity and image problems, but those are the journal's contracted systems, not public chatbots, and flagged findings still require human verification. ### What if I remove the author names before uploading the manuscript to an AI tool? That does not make it allowed. The confidentiality obligation covers the unpublished findings and the whole manuscript, not just identifying details. ### Can AI be listed as a reviewer or as an author? No. COPE and ICMJE guidance state that AI tools cannot hold authorship or reviewer status because accountability and responsibility can only rest with humans. ### My journal has no explicit AI policy. What should I do? Follow the underlying confidentiality and accountability rules that already govern peer review, use only language assistance on your own text, and ask the editor directly about anything borderline. ### Are AI-detection tools reliable for spotting AI-written reviews? No. Detection tools produce both false positives and false negatives, which is why journals rely on policy, disclosure, and human editorial judgment rather than automated detection. ## Further Reading - [Prevalence, determinants, and consequences of workplace violence in emergency departments - a systematic review.](https://pubmed.ncbi.nlm.nih.gov/43000232/) ## Sources 1. [When Forensic Pathology Fails: Ethical Lessons From International Regulatory, Judicial, and Historical Human-Rights Cases.](https://pubmed.ncbi.nlm.nih.gov/42992540/) 2. [Peer victimization and family dynamics in the association between childhood abuse and adolescent non-suicidal self-injury: a cross-sectional moderated mediation study.](https://pubmed.ncbi.nlm.nih.gov/42983887/) 3. [Explainable AI-based techniques for disease classification: a systematic review and future work.](https://pubmed.ncbi.nlm.nih.gov/42989268/) 4. [Distinguishing artificial intelligence from robotic execution in total knee arthroplasty: a narrative review and evidence map.](https://pubmed.ncbi.nlm.nih.gov/42989357/) 5. [A scoping review of autoethnographic approaches in medical education research.](https://pubmed.ncbi.nlm.nih.gov/43006209/)