01 · The Question
Can a Reviewer Give a Confidential Manuscript to AI?
You agree to review a manuscript. The paper is dense, the methods are unfamiliar in places, and the deadline is close. An AI system could summarize the manuscript, identify possible methodological weaknesses, check the statistics, compare the argument with existing literature, or draft a review.
Technically, all you need to do is upload the PDF.
But the manuscript was not given to you for unrestricted use. You received it because a journal entrusted you with unpublished material for the specific purpose of confidential peer review. That relationship substantially changes what you may do with the file.
03 · What You Need to Know
Peer Review Gives You Access, Not General Permission to Redistribute
The manuscript is entrusted to you for a limited purpose
Peer review depends on researchers being willing to share unpublished work with editors and reviewers before publication. Confidentiality is part of the mechanism that makes this possible.
ICMJE describes manuscripts submitted to journals as privileged communications and authors' private, confidential property. It states that reviewers must keep manuscripts and the information they contain strictly confidential, must not publicly discuss the work, and must not appropriate authors' ideas before publication.
Elsevier similarly requires manuscripts received for review to be treated as confidential documents and prohibits reviewers from sharing review information without appropriate permission.
Uploading a manuscript to an external AI service is therefore not merely "using software." It can constitute disclosure of material that was entrusted to you under confidentiality.
The restriction can cover information from the manuscript, not only the PDF
Confidentiality does not vanish when information is copied out of the original file.
Consider these prompts:
- "Summarize this unpublished manuscript," followed by the full text.
- "Is this result novel?" followed by the paper's central unpublished finding.
- "Find flaws in this method," followed by three paragraphs from the Methods section.
- "Help me write my review," followed by your notes describing the paper's results and weaknesses.
In each case, the AI may receive confidential information even though only the first prompt includes the complete manuscript.
Elsevier's current reviewer policy makes this explicit: reviewers must not upload a submitted manuscript, any part of it, or information from it into AI tools. The concern includes authors' confidentiality, proprietary rights, and potential data-privacy rights.
Different journals can impose different AI rules
There is no single worldwide peer-review AI policy.
ICMJE takes a principles-based approach. Reviewers must follow the journal's stated AI policy or request permission before using AI to facilitate their review. It states that maintaining manuscript confidentiality may prohibit uploading the manuscript to AI technology where confidentiality cannot be assured unless the journal explicitly permits such use.
Elsevier's current journal policy is more specific. Reviewers must not upload submitted manuscripts or information from them into AI tools. It allows AI only in a supportive capacity, such as improving the language or structure of a review report or assisting with background literature searches, provided confidentiality is maintained, human control remains, and an appropriate private AI tool is used. Qualifying AI assistance must also be disclosed.
The practical rule is therefore straightforward: the journal's actual policy controls your review assignment.
| Proposed AI use |
Confidentiality issue |
What to do |
| Upload the entire manuscript for summarization |
Discloses the confidential manuscript |
Do not do this unless explicitly permitted under the journal's applicable policy and confidentiality requirements |
| Paste a methods section for criticism |
Discloses part of the confidential manuscript |
Treat manuscript excerpts as confidential too |
| Describe unpublished findings without quoting them |
Can still disclose information obtained through review |
Do not assume paraphrasing removes confidentiality |
| Ask AI about a generic statistical method without mentioning the paper |
May require no manuscript information |
Potentially possible if consistent with journal policy |
| Improve the grammar of your review report |
The report itself may reveal manuscript information |
Follow the journal's specific policy and remove confidential content where necessary |
"I did not upload the manuscript" is not always enough
A reviewer can disclose confidential research without attaching a file. Detailed prompts can reveal the manuscript's hypotheses, results, methods, limitations, identities, or novel ideas.
This is particularly important when using AI to fact-check or critique the manuscript. To obtain a highly specific answer, the reviewer may be tempted to provide precisely the unpublished details that confidentiality protects.
The correct unit of analysis is therefore information disclosed, not simply whether a PDF was uploaded.
A private or enterprise AI tool does not automatically override journal policy
Data protections matter. Elsevier currently defines a private AI tool for reviewer support as one that does not retain, reuse, share, or learn from submitted content beyond what is necessary to provide the service. It recommends verifying privacy policies and terms rather than assuming that a paid or enterprise product is private.
Even so, technical privacy and editorial permission are different requirements. A highly secure institutional AI system does not grant you permission to upload a manuscript when the journal prohibits it.
Technical confidentiality
Concerns how the AI system stores, uses, shares, protects, and provides access to submitted information.
Editorial authorization
Concerns whether the journal permits you to provide peer-review information to that AI system at all.
You may need both. One does not substitute for the other.
Turning off model training is not sufficient
A reviewer might reason that the manuscript is safe because the AI service promises not to use prompts for model training. That addresses one possible use of the information but not the underlying disclosure.
The manuscript still leaves the review environment and is processed by another system. Retention, access, logging, service providers, contractual terms, and journal policy can remain relevant. More fundamentally, a publisher may prohibit the upload regardless of training settings.
This is why whether prompts are used for AI model training cannot answer the peer-review confidentiality question by itself.
Peer-review confidentiality also protects intellectual priority
A manuscript may contain ideas, results, methods, inventions, or interpretations that have not yet entered the public record. Reviewers gain early access because the journal needs expert evaluation, not because the material has become available for general use.
ICMJE specifically states that reviewers must not appropriate authors' ideas before publication. Elsevier likewise states that unpublished materials disclosed in a submitted manuscript must not be used in the reviewer's own research without the author's express written consent.
AI disclosure can therefore implicate not only privacy but also authors' proprietary interests and expectations concerning intellectual priority.
Peer-review material may contain personal or sensitive information
Some manuscripts contain participant quotations, case histories, clinical details, photographs, geographic information, or other potentially identifiable data. Uploading such material can add a data-privacy issue to the existing editorial confidentiality problem.
Elsevier's policy explicitly identifies possible breaches of personally identifiable information as one reason reviewers must not upload manuscripts into AI tools.
Even a manuscript that has passed a journal's anonymization process may contain information that remains sensitive or indirectly identifiable.
Review reports and editorial correspondence can also be confidential
Confidential peer-review material extends beyond the author's PDF. It can include reviewer reports, confidential comments to the editor, editor correspondence, decision letters, supplementary files, reviewer identities, manuscript-status information, and notes created from the manuscript.
ICMJE states that editors should not disclose manuscript content, review status, reviewer criticism, or the ultimate fate of a manuscript beyond appropriate participants in the process. Elsevier's policies similarly extend confidentiality to review-related communications.
Therefore, using reviewer comments with AI requires attention to what those comments reveal and the role in which you received them.
AI also creates an assessment-quality problem
Even if confidentiality were fully addressed, outsourcing the intellectual assessment of a manuscript raises a separate problem. Peer review requires disciplinary judgment for which the invited reviewer is accountable.
ICMJE warns that AI can generate output that sounds authoritative while being incorrect, incomplete, or biased. Elsevier states that AI cannot replace reviewers' critical thinking or independent evaluation and that reviewers remain responsible for their reports.
A reviewer was invited because of their expertise. Sending the paper to a model and asking, "What is wrong with this?" rather misses the job description.
Watch Out
Do not assume that paraphrasing an unpublished result, copying only one paragraph, or using a no-training AI account eliminates peer-review confidentiality. The question is whether confidential information is being disclosed and whether the journal permits that disclosure.
07 · A Quick Checklist
Before Using AI During Confidential Peer Review
Before involving an AI tool in a review, check:
Read the specific journal or publisher's current AI policy for reviewers.
Treat the manuscript, supplementary files, unpublished findings, review notes, and information obtained from the manuscript as confidential.
Do not assume that an excerpt, paraphrase, screenshot, table, figure, or detailed description falls outside the confidentiality obligation.
Do not infer permission from a no-training setting, private mode, paid account, or enterprise label.
Use public sources or generic questions when you need background information that can be obtained without revealing manuscript content.
If limited AI assistance is permitted, verify that the tool meets the journal's confidentiality and privacy requirements.
Keep the scientific evaluation and final judgment under your own control and verify any AI-assisted material independently.
Disclose permitted AI use to the journal when its policy or applicable publishing guidance requires disclosure.
