1. Editorial purpose and boundaries of AI assistance
Artificial intelligence can support selected editorial tasks, but it must not displace the independent judgment of editors or the obligations owed to authors, reviewers and readers. An appropriate use case begins with a clearly defined editorial problem, such as organizing administrative checks, discovering potential reviewers or identifying passages that warrant closer human examination.
Editors should distinguish administrative automation, conventional analytical software and generative systems that produce new text or interpretations. Their risks differ. A deterministic formatting check is not equivalent to a model that generates a scientific critique from an unpublished manuscript, and the approval process should reflect those differences.
Editorial responsibility remains with the journal even when a vendor, platform or publisher supplies the tool. No automated output should become an acceptance, rejection, integrity finding or reviewer-selection decision without appropriate human evaluation.
2. Classify tools and approve specific use cases
Maintain an inventory of tools proposed for editorial use, including their purpose, provider, data inputs, outputs, integration points and authorized users. Approval should be tied to a particular workflow and data category rather than granted broadly to any product marketed as an editorial assistant.
Distinguish tools operating entirely inside an approved publishing environment from external services that receive manuscript content or metadata. Assess whether the provider stores prompts, retains documents, uses inputs for model training or passes information to subprocessors. Vendor marketing statements are not a substitute for documented terms and technical review.
Where a tool changes model, provider, retention practice or intended use, reassess its approval. A previously acceptable metadata classifier may require a new review before it is allowed to process full manuscript files or confidential reviewer reports.
3. Apply confidentiality before testing or uploading
Unpublished manuscripts, reviewer reports, author identities, editorial correspondence and research-integrity allegations may be confidential. Editors must not paste or upload this material into an unapproved AI service, even when the intended task appears routine or the service is widely used.
Before a tool receives any information, identify the minimum data necessary, the permitted purpose, who can access the output and whether the workflow complies with the journal’s review model and applicable publisher requirements. Redaction can reduce exposure but should not be assumed to make sensitive research or personal information anonymous.
Test new tools using synthetic, public or expressly authorized material. If a suspected disclosure occurs, stop the relevant use, preserve the necessary incident record and follow the publisher’s confidentiality and data-incident procedures.
4. Evaluate privacy, data governance and vendor terms
Document what data the system collects, where it is processed, how long it is retained and whether it may be reused to improve models. Consider access controls, deletion mechanisms, subprocessors, geographic transfers and the provider’s procedures for security incidents.
Manuscripts can contain unpublished intellectual property, identifiable participant information, proprietary methods or confidential commercial details. Data minimization and contractual safeguards must therefore be evaluated against the actual material a workflow handles, not merely the tool’s advertised purpose.
Approval should identify who may authorize a new integration and how compliance is reviewed. Editors should not individually accept conflicting external terms on behalf of the journal or assume that a personal subscription has the protections of an institutional agreement.
5. Keep humans accountable for all consequential decisions
AI-generated summaries, risk flags, reviewer suggestions and draft correspondence are aids for human consideration, not independent evidence of scientific quality or misconduct. A responsible editor should be able to explain the decision using the manuscript, applicable policies and verifiable editorial evidence without relying on an opaque model score.
Establish review points where an authorized person checks the input, verifies material claims, corrects errors and determines whether the output is suitable for the next step. Higher-impact uses, including integrity triage or editorial recommendations, require stronger scrutiny than routine administrative assistance.
Do not create a workflow in which automation silently determines which manuscripts are rejected, which authors receive additional scrutiny or which reviewers are excluded. Accountability includes the ability to challenge, override and audit automated suggestions.
6. Use AI for submission checks without confusing flags with findings
Approved tools may help identify missing declarations, inconsistent metadata, incomplete files, broken references or other administrative issues. Their output should guide targeted verification rather than automatically become a statement that the submission violates policy.
Automated checks may fail on unusual article types, multilingual material, interdisciplinary work or nonstandard reporting formats. Editorial staff should be able to correct false flags and distinguish mandatory requirements from journal preferences.
Communicate verified deficiencies in clear terms. Authors should not be asked to respond to unexplained model labels or speculative assertions that the editor has not independently assessed.
7. Support reviewer discovery while verifying identity and independence
Automated matching can suggest researchers whose publications appear relevant to a manuscript. The editor must still assess the candidate’s actual expertise, verify a credible contact route and screen for competing interests and relevant professional relationships.
Recommendation systems may overrepresent frequently indexed authors, established institutions, English-language publications or familiar collaboration networks. Search beyond the top-ranked results when necessary to construct an expert and sufficiently independent panel.
Do not submit confidential full-text manuscripts to unapproved matching services. If an approved system uses manuscript content, its data processing and access arrangements should be documented and compatible with the journal’s confidentiality obligations.
8. Assess similarity, citations and publication overlap responsibly
Similarity-detection tools can locate matching text or related publications, but a similarity score is not a plagiarism finding. Editors should inspect the location, context, attribution, extent and scholarly significance of overlap before taking action.
Automated citation and reference checks may flag fabricated-looking references, missing identifiers or unusual citation patterns. Confirm the actual source and its relevance before contacting authors; bibliographic databases and AI-generated metadata can themselves contain errors.
Preprints, legitimate methods reuse, quotations and related publications require contextual assessment under the journal’s policies. An AI-generated accusation must never replace the evidence-based process used for research-integrity concerns.
9. Treat image, data and statistical alerts as investigative leads
Analytical systems may identify duplicated image regions, inconsistent numerical reporting or unusual patterns in datasets. Their outputs should be treated as leads for expert verification, not as conclusive evidence of fabrication or falsification.
False positives may arise from legitimate image reuse, shared controls, compression artifacts, reporting conventions or characteristics of the study design. Editors should document the exact concern and seek original material or specialist advice where proportionate.
Do not upload identifiable participant data, unpublished datasets or confidential figures to an external system without explicit authorization and suitable safeguards. Complex forensic questions may exceed a journal’s technical authority and require institutional involvement.
10. Understand the limits of AI-generated scientific summaries
Generative tools can omit qualifying details, misstate methods, invent references and express unwarranted confidence. A fluent summary is not evidence that the system understood the manuscript, and errors may be especially consequential when the summary informs reviewer selection or editorial triage.
If an approved tool produces a summary, an editor should verify material descriptions against the source manuscript and should not rely on it to determine novelty, methodological validity or ethical compliance. Reviewers must still receive the information necessary to make their own independent assessment.
Outputs should identify their role as internal assistance where appropriate. Avoid passing generated critiques to authors as if they were independently produced peer-review reports.
11. Draft editorial correspondence with appropriate verification
An approved drafting tool may help improve clarity, organization or language in routine communications. The editor must check every substantive statement, manuscript-specific instruction, policy reference and representation of reviewer concerns before sending the message.
Decision letters should reflect actual editorial reasoning and distinguish essential revisions from optional suggestions. A model must not invent reviewer positions, promise acceptance, introduce new conditions without editorial approval or produce accusatory language from uncertain integrity signals.
Use approved workflows and the minimum necessary information. A convenient external chatbot is not an acceptable destination for confidential reports or unpublished findings simply because the final message is intended for the authors.
12. Govern AI use in peer review and co-reviewing
Reviewers should follow the journal’s confidentiality and AI-use requirements. They should not place unpublished manuscripts, figures or reports into unapproved generative systems, nor delegate independent scholarly judgment to a model.
Where journal policy permits limited assistance, define what may be used, what disclosure is required and who remains responsible for the accuracy and confidentiality of the report. AI-assisted language editing differs from asking a model to generate the scientific assessment itself.
Editors should handle suspected noncompliance through a fair, evidence-based process. Unusual writing style alone does not establish that a reviewer used AI or breached confidentiality.
13. Avoid unreliable AI-detection claims and fabricated evidence
AI-text detectors can produce false positives and false negatives and should not be treated as definitive proof that a manuscript or review was AI-generated. Language background, editing practices and disciplinary conventions can influence detector behavior.
Where a concern matters to publication policy, focus on verifiable issues such as inaccurate disclosure, invented citations, unsupported claims, manipulated images or documented confidentiality breaches. Request specific explanations and evidence rather than confronting authors with an unexplained detection percentage.
Editors should record the limitations of any automated evidence and avoid making misconduct findings from a single opaque score. Serious allegations belong in the journal’s established integrity procedure.
14. Test fairness, accessibility and performance across disciplines
AI systems may perform unevenly across languages, research fields, article types, career stages and institutional contexts. Before deployment, test representative examples and inspect whether error rates or recommendations systematically disadvantage particular categories of submissions.
Provide a route for editorial override and correction. Authors should not be disadvantaged because an automated checker cannot parse an accessible file format, a specialized method or a legitimate nonstandard manuscript structure.
Monitor the system after deployment. A tool that performs adequately on routine research articles may behave differently with case reports, qualitative studies, registered reports or interdisciplinary submissions.
15. Document tool use and preserve an auditable editorial record
Maintain proportionate records of which approved tool was used, for what task, by whom and with what consequential output. The journal should be able to reconstruct material decisions from verifiable evidence and human reasoning, not from a model recommendation that cannot be explained.
Retention should respect confidentiality and data-protection requirements. Do not preserve unnecessary full prompts or duplicate confidential manuscript files merely to demonstrate that a tool was used; record the minimum information needed for oversight.
When an automated suggestion materially influences reviewer selection, integrity triage or another consequential step, document the human verification and resulting editorial rationale.
16. Train editors and establish operational safeguards
Approval alone is insufficient if staff do not understand a tool’s permitted scope. Provide clear guidance on approved services, prohibited uploads, verification requirements, disclosure expectations, incident reporting and the distinction between assistance and delegated judgment.
Access should follow editorial roles and be removed when no longer needed. New features should not become available to all users automatically when they introduce additional data sharing or more consequential automated recommendations.
Regularly review recurring errors, unexpected outputs and user workarounds. A practical governance program should make compliant behavior straightforward rather than relying on individual memory of complex rules.
17. Respond to AI-related incidents and policy breaches
If confidential material may have been exposed, stop the relevant workflow, restrict further access and use the publisher’s established incident-response procedure. Preserve enough information to determine what was submitted, where it went and which parties may be affected.
An unreliable automated flag or misleading generated decision letter may require correction, independent reassessment or communication with affected parties. Match the remedy to the actual harm or risk rather than treating every tool error as misconduct.
Investigate suspected policy violations fairly. Distinguish inadvertent misuse, unclear training, tool malfunction and deliberate circumvention, and refer serious integrity or confidentiality matters to the appropriate authorized process.
18. A repeatable approval and use framework
Before using an AI editorial tool, define the exact task and data involved, check that the service is approved, assess confidentiality and vendor terms, and determine what independent verification the output requires. If any of these conditions is unresolved, do not upload confidential material.
During use, minimize inputs, verify substantive claims, document consequential recommendations and ensure that an authorized human retains control. Where outputs are uncertain, obtain appropriate expertise rather than treating the model’s confidence as evidence.
After deployment, review accuracy, fairness, security incidents and changes in provider behavior. The governing principle is useful assistance under accountable editorial control—not automation for its own sake.