03 · What You Need to Know
How AI Processing Can Conflict With Peer-Review Confidentiality
What Does Peer-Review Confidentiality Actually Protect?
Peer-review confidentiality protects information entrusted to reviewers through the journal's editorial process. The obligation generally covers more than the manuscript document itself.
It may include unpublished research findings, methodological details, supplementary files, editorial correspondence, reviewer identities, confidential assessments, and information about the manuscript's status.
The International Committee of Medical Journal Editors (ICMJE) describes submitted manuscripts as privileged communications and emphasizes that reviewers must keep the manuscripts and their contents confidential.
The Committee on Publication Ethics (COPE) likewise identifies confidentiality as a fundamental responsibility of peer reviewers.
These principles help protect authors from premature disclosure of their work and maintain trust in the editorial process.
Why AI Processing Can Constitute a Disclosure
When a reviewer submits manuscript information to an external generative AI service, the information may be transmitted to systems operated by another organization.
Depending on the service, processing may involve remote infrastructure, retention of submitted content, authorized personnel access, or subprocessors.
These arrangements do not necessarily mean the information becomes publicly available. However, confidentiality obligations can restrict disclosure to unauthorized recipients even when the material is not published openly.
The relevant question is whether the reviewer was authorized to transmit the information to that particular processing environment.
For example, a journal may authorize a reviewer to access a manuscript through its editorial platform without authorizing the reviewer to submit its contents to a third-party AI service.
That distinction applies regardless of whether the AI generates a useful or scientifically accurate response.
Confidentiality Is Not Limited to Uploading a PDF
A common misunderstanding is that confidentiality is breached only when the complete manuscript file is uploaded.
In reality, the content of a prompt may itself disclose protected information.
| Information Shared With AI |
Potential Confidentiality Concern |
| Complete submitted manuscript |
Transfers the unpublished submission to another processing environment |
| Methods section or research protocol |
May disclose unpublished procedures or intellectual contributions |
| Results table or figure |
May reveal unpublished findings or sensitive data |
| Detailed manuscript summary |
May reproduce confidential scientific information in paraphrased form |
| Reviewer comments |
May reveal manuscript content, confidential assessment, or editorial information |
| Generic methodological question |
May avoid manuscript disclosure if no confidential details are included |
The size of the material does not determine whether disclosure is authorized. A short paragraph describing a novel technique may be more sensitive than several pages of publicly established background information.
Reviewers should therefore consider the information conveyed by an AI prompt, not merely whether a document was attached.
What Do Current Publisher Policies Say?
Publisher policies differ, and their requirements may change. Reviewers should consult the current policy governing the specific manuscript.
The ICMJE's recommendations on AI use by reviewers require reviewers to follow the journal's stated policy or request permission before using AI to facilitate their review. The guidance emphasizes confidentiality and disclosure of AI assistance.
Elsevier's generative AI policies for journals, updated in June 2026, explicitly prohibit reviewers from uploading submitted manuscripts or information from them into AI tools. The policy permits certain supportive uses, such as improving review-report language or conducting background literature searches, provided confidentiality is maintained and the specified requirements are met.
These are not interchangeable policies. The ICMJE provides recommendations that journals may adopt, whereas Elsevier establishes requirements for its covered journal activities.
A reviewer should not assume that permission granted by one publisher applies to another.
General publication-ethics guidance
Establishes principles such as confidentiality, accountability, and transparency.
Publisher or journal policy
Specifies what is permitted or prohibited within a particular editorial process.
Does Author Permission Resolve the Confidentiality Problem?
Author permission may be relevant, but it is not necessarily sufficient on its own.
The ICMJE's general guidance addresses the need for authors' explicit permission when manuscript material would be uploaded to systems where confidentiality cannot be assured. Its reviewer-specific guidance also requires compliance with the journal's policy or permission from the journal.
A publisher may impose a stricter prohibition that does not allow an individual reviewer to authorize an exception.
Consequently, reviewers should not independently approach authors or assume that informal consent overrides the journal's editorial procedures.
Any proposed exception should be handled through the appropriate editorial authority and should satisfy applicable confidentiality and data protection requirements.
Does a Private or Enterprise AI System Eliminate the Risk?
No. A private processing arrangement may reduce certain risks, but it does not automatically establish that the proposed use is authorized.
AI services differ in their data retention, training, access control, and contractual arrangements. Some institutional or enterprise services may provide stronger protections than ordinary consumer interfaces.
However, the journal may still prohibit sharing unpublished manuscript information with such systems.
Elsevier's current reviewer guidance illustrates this distinction: although it discusses private AI tools for permitted supportive activities, it continues to prohibit reviewers from uploading confidential manuscript information.
Reviewers should therefore distinguish technical privacy protections from permission to process confidential editorial material.
Does Removing Author Names or Identifying Details Make AI Use Acceptable?
Not automatically.
Removing author names may reduce identification risks, but the manuscript's scientific content can remain confidential.
Distinctive datasets, institutional contexts, experimental designs, or unpublished findings may also make the work recognizable.
Furthermore, peer-review confidentiality protects more than personal identity. It protects information disclosed to the reviewer for a limited editorial purpose.
A blinded manuscript is therefore not equivalent to a publicly released document.
What If the Reviewer Uses AI Only to Improve Their Own Comments?
AI-assisted language editing may be permissible under some journal policies, but the content of the comments still matters.
For example, a reviewer might write a detailed criticism describing an unpublished experimental procedure and paste it into an AI tool for rewriting. Although the text was written by the reviewer, it may disclose confidential manuscript information.
By contrast, requesting a generic explanation of how to make academic feedback more constructive may not require sharing anything about the assigned submission.
The distinction concerns both the journal's permission and the information transmitted.
Reviewers should follow the applicable rules for AI-assisted improvement of reviewer comments without assuming that editing their own words removes confidentiality obligations.
Can AI Be Used to Check a General Methodological Principle?
Potentially, if the journal permits the intended AI use and the question does not disclose confidential manuscript information.
A reviewer might seek a general explanation of assumptions underlying logistic regression, common threats to causal inference in observational studies, or the interpretation of confidence intervals.
Such questions can often be formulated without describing the manuscript under review.
However, the AI's explanation must still be verified against appropriate methodological references. Generative AI may produce incorrect or incomplete technical advice.
Using AI to clarify a general concept is different from asking AI to evaluate the confidential manuscript's methodological weaknesses.
Neither activity should be undertaken contrary to the journal's policy.
What About Publicly Available Preprints?
Some manuscripts submitted to journals are also available as public preprints.
The public version may be accessible to anyone, but the journal submission remains part of a confidential editorial process.
The submitted version may differ from the preprint, contain additional unpublished material, or be accompanied by confidential correspondence.
Moreover, the journal may restrict AI use in peer review regardless of the public availability of related material.
Reviewers should therefore distinguish information independently available in the public domain from information obtained through their confidential review assignment.
Can AI Use Violate Confidentiality Without Any Visible Data Leak?
Yes, depending on the applicable confidentiality obligations.
A breach of confidentiality does not necessarily require proof that information was publicly exposed, used for training, or accessed by an unauthorized individual.
Unauthorized transmission or processing may itself violate the terms under which the reviewer received the material.
This distinction is important because reviewers may mistakenly conclude that no confidentiality problem exists if an AI service produces a response privately and the conversation is never shared.
The absence of a visible leak does not establish that the processing was authorized.
What Should Reviewers Do After a Possible Unauthorized Disclosure?
If a reviewer believes confidential manuscript information has been submitted to an AI service without permission, the appropriate response is to stop further processing and seek guidance promptly.
The handling editor or designated journal contact should ordinarily be informed. Depending on the material and circumstances, institutional privacy or information-security personnel may also need to be involved.
The reviewer should provide an accurate account of what was submitted, which service received it, and when the interaction occurred.
Deleting the visible conversation may be useful under an applicable incident-response procedure, but reviewers should not assume that deletion guarantees the removal of all retained information.
The journal and relevant institutional authorities can determine what additional steps are required.
Why Confidentiality and Scientific Responsibility Are Separate Obligations
An AI system might generate an accurate methodological criticism while still having received the manuscript through an unauthorized disclosure.
Conversely, a reviewer might use an approved system but receive an incorrect or irrelevant critique.
These are separate concerns. Confidentiality addresses whether the information was processed appropriately. Scientific responsibility concerns whether the assessment is accurate, fair, and independently defensible.
Both must be satisfied.
The broader question of responsible AI assistance in peer-review writing therefore cannot be resolved solely by evaluating the quality of the generated text.
