Purpose and scope of responsible AI use
AI tools can assist with narrow tasks, but peer review remains an independent expert assessment entrusted to a named reviewer. This guide covers generative systems, automated writing aids, translation, search assistants, code assistants, image-analysis software and other tools that process review material.
The key distinction is not whether a tool is marketed as AI. It is whether the tool receives confidential material, produces substantive judgments, stores data, or influences the reviewer’s report. Reviewers should examine those functions before use.
Journal-specific instructions take precedence where they are more restrictive. No reviewer should infer permission merely because a tool is publicly available or commonly used.
The reviewer remains personally accountable
The invited reviewer is responsible for reading the manuscript, evaluating the evidence and writing a defensible report. An automated system cannot accept that professional responsibility or independently resolve an editorial judgment.
Any permitted tool output must be treated as a draft, suggestion or technical aid. The reviewer should verify every substantive criticism against the manuscript and their own expertise, especially when a system produces confident explanations without traceable evidence.
A report should not contain claims that the reviewer cannot explain or defend if the editor asks for clarification.
Read the journal’s AI policy before opening a tool
Reviewers should consult the journal’s reviewer instructions and any applicable publisher policy before placing manuscript content into a digital service. A policy may prohibit external AI processing, permit narrowly defined uses, or require prior editorial approval and disclosure.
When the policy is silent or ambiguous, the reviewer should ask the editor rather than assuming permission. Approval for one task, such as language editing, does not automatically authorize summarization of an entire unpublished manuscript.
The reviewer should document any authorization that materially affects confidentiality or the review process.
Confidentiality applies to prompts and outputs
Unpublished titles, abstracts, methods, figures, datasets, reviewer invitations and editorial correspondence can all be confidential. Pasting even a short passage into an external prompt may disclose protected material.
Removing author names is not sufficient if the remaining methods, dataset, study setting or distinctive results can identify the work. A generated output can also reproduce confidential information and should be handled accordingly.
The safe default is to avoid uploading review material to an external AI service unless the journal has explicitly authorized the arrangement and its data-handling conditions.
Understand storage, training and onward disclosure
Before using an authorized service, determine what information is transmitted, where it is stored, how long it is retained, whether it may be used to train models, and whether subcontractors or human reviewers may access it.
Marketing descriptions such as private, secure or enterprise-grade do not by themselves establish compatibility with the journal’s confidentiality requirements. Actual contractual terms, configuration and institutional safeguards matter.
If those conditions cannot be established, use a workflow that does not disclose manuscript content or consult the editor.
Distinguish low-risk assistance from manuscript processing
General assistance that does not expose confidential material—such as asking a tool to explain a public statistical concept or generate a blank report outline—raises different risks from uploading a manuscript for summary or critique.
Risk increases when a service receives full text, figures, patient information, supplementary datasets, editorial correspondence or unpublished results. Even a seemingly harmless translation request may transmit protected content.
Classify the proposed task and the information involved before deciding whether it is permitted.
AI must not replace critical reading
A model-generated summary may omit eligibility criteria, misstate endpoints, confuse denominators or overlook limitations that are central to a scientific judgment. Reviewers should read the original manuscript and supporting material themselves.
If a journal authorizes a tool to assist with navigation or organization, compare its output with the source and do not let the generated framing determine which findings receive attention.
The review should reflect the reviewer’s independent assessment rather than a reformatted model response.
Hallucinated citations and fabricated methodological concerns
Generative systems can invent papers, quotations, reporting requirements, statistical flaws and technical explanations. Their fluency does not establish accuracy.
Every citation suggested for a reviewer report should be checked against an authentic source and assessed for genuine relevance. Every methodological criticism should be tied to the actual design, analysis or reported evidence.
Do not pass unsupported AI-generated allegations to the editor or authors. If a claim cannot be verified, remove it or identify the uncertainty.
Statistical and computational assistance
Tools may help a qualified reviewer reason about code syntax, a public statistical formula or a non-confidential toy example. They should not be treated as substitutes for specialist statistical review.
If journal policy permits analysis of supplied code or data, protect the files under the same confidentiality and access controls as the manuscript. Confirm software versions, assumptions, inputs and outputs independently.
When the central method lies outside the reviewer’s competence, request a specialist assessment rather than using AI to simulate expertise.
Language editing and translation
A reviewer may need help expressing comments clearly or understanding unfamiliar terminology. Use of language tools should follow the same confidentiality rules as any other external processor.
A safer approach is to edit reviewer-authored generic language that contains no manuscript-specific information, provided the journal allows it. Translation of unpublished passages or complete reviewer reports can disclose confidential material.
After any permitted language assistance, reread the text to ensure that criticism has not become harsher, less precise or scientifically altered.
Literature discovery and citation checking
AI-assisted search may help identify publicly available background literature, but search results and generated bibliographies require independent verification.
Do not enter unpublished hypotheses or distinctive manuscript details into external discovery tools unless authorized. Prefer neutral public search terms when researching background knowledge.
Suggested citations should correct genuine omissions or support a specific scientific concern; they should never become coercive citation requests.
Image and figure assessment
Automated image tools can produce false positives and false negatives. Compression, legitimate processing and experimental context may affect apparent anomalies.
Reviewers should describe concrete observations and, where appropriate, alert the editor confidentially. They should not upload unpublished figures to an unauthorized image-analysis service or present an automated flag as proof of manipulation.
Editors should coordinate any formal image-integrity investigation under the journal’s established procedures.
Patient, participant and sensitive research information
Clinical images, genetic data, small-cell tables, qualitative quotations and location-specific datasets may remain identifiable even after obvious identifiers are removed.
External AI processing can create additional privacy, consent, legal and institutional risks. Reviewers should not attempt to reidentify participants or upload sensitive materials for convenience.
If evaluation requires specialist access or verification, request an editorially managed process with appropriate safeguards.
Bias, framing and unequal treatment
AI outputs may reflect training-data gaps, disciplinary conventions, language biases or unwarranted assumptions about authors and institutions. Such output should not determine whether a manuscript receives fair consideration.
Evaluate claims according to design, evidence and reporting. Do not use automated predictions about author quality, institution prestige, nationality or writing style as substitutes for scientific assessment.
Reviewers should check whether permitted assistance changes the tone or evidentiary standard applied to the work.
AI detection is not evidence of misconduct
Automated detectors of AI-generated writing can be unreliable, particularly across languages, writing styles and edited text. A score or label is not a finding of undisclosed AI use or research misconduct.
If the manuscript contains a concrete inconsistency, fabricated reference or other verifiable problem, document that problem itself. Do not accuse authors solely because prose resembles model-generated text.
Route material integrity concerns through the editor and distinguish observed evidence from speculation.
Authorship, originality and undisclosed AI in the manuscript
Reviewers may assess whether methods, citations, images or disclosure statements are sufficiently transparent under the journal’s author policies. The relevant question is whether a specific issue affects reliability, accountability or policy compliance.
Do not presume that any use of AI by authors is prohibited. Policies can distinguish language assistance from AI-generated data, images, analyses or substantive writing.
When disclosure appears inadequate, identify the precise information needed and allow the editor to apply the journal’s requirements.
Co-reviewing and automated delegation
Permission to involve a trainee or specialist colleague is separate from permission to use an AI service. Neither form of assistance should occur secretly.
A reviewer should not delegate the entire scientific assessment to a tool and present the result as personal expert work. Any authorized human contributor must understand confidentiality and be identified as required by the journal.
The invited reviewer remains responsible for the accuracy and completeness of the final report.
Drafting the report with permitted assistance
Start from the reviewer’s own notes, observations and prioritized concerns. If limited AI-assisted drafting is authorized, provide only information that the journal permits the tool to process.
Review the resulting language line by line. Remove invented manuscript details, unsupported conclusions, repetitive generic requests and any shift in recommendation that does not follow from the evidence.
The final report should clearly separate major scientific issues, minor reporting issues and confidential information intended only for the editor.
Disclose AI assistance accurately
Follow the journal’s specific disclosure rules. A useful disclosure identifies the kind of tool and the task it performed without unnecessarily reproducing confidential prompts or manuscript information.
Disclosure is not a cure for an otherwise prohibited upload or breach. Permission and confidentiality must be addressed before the tool is used.
Keep a proportionate internal record of authorized assistance if needed to answer legitimate editorial questions.
When AI use creates an error or breach
If a reviewer discovers that a tool introduced a false claim, incorrect citation or distorted interpretation, correct the report promptly and notify the editor when the error may have affected editorial assessment.
If confidential material was uploaded to an unauthorized service or disclosed through another channel, stop further processing and inform the editor without delay. Follow institutional and journal incident procedures rather than trying to conceal the event.
Preserve only the information necessary for authorized incident assessment and do not spread the confidential material further while investigating.
AI use during revision and re-review
When authors submit a revision, the reviewer should evaluate the revised manuscript and response against the original concerns. Automated comparison may miss substantive changes or invent differences.
Do not introduce new demands solely because an AI tool suggests them. A second-round review should focus on whether material concerns were addressed and on genuinely new problems introduced by revision.
The same confidentiality, authorization and disclosure rules apply to every review round.
A practical risk assessment before each use
Before using any AI-assisted tool, ask whether the task is permitted, whether it requires confidential inputs, whether the service stores or trains on those inputs, whether sensitive participant data are involved, and whether the output can be independently verified.
Also consider whether the tool adds real value. If the same task can be completed reliably without disclosing manuscript content, choose the lower-exposure workflow.
Where any material condition remains unresolved, do not upload the material; ask the editor for guidance.
The Treata AI peer-review standard
Treata’s reviewer guidance places human accountability, independent scientific judgment and manuscript confidentiality above convenience. AI assistance is conditional on applicable journal permission, suitable information safeguards and any required disclosure.
Reviewers must verify substantive outputs, avoid fabricated criticism and citations, protect participant data, and report any breach or integrity concern through the editor.
Journal-specific rules may be stricter. This guidance should be read alongside the publisher’s Peer Review, Artificial Intelligence, Competing Interests and Research Integrity policies.
Related reviewer guides
Continue reading
Canonical Treata standards