Generative AI and AI-Assisted Technologies Policy

This policy may apply to tools used for:

• text generation;
• language editing;
• summarisation;
• translation;
• reference generation;
• data analysis;
• statistical assistance;
• coding;
• image generation or modification;
• literature screening;
• information extraction;
• machine-learning model development;
• peer-review assistance; or
• other automated scholarly or editorial tasks.

The requirements depend on how materially the tool contributes to the work and whether its use affects transparency, confidentiality, reproducibility, authorship, evidence, or scientific interpretation.

Authorship requires human responsibilities that AI systems cannot fulfil, including the ability to:

• take responsibility for the accuracy of the work;
• approve the final manuscript;
• disclose competing interests;
• respond to editorial queries;
• defend the integrity of the work;
• approve corrections or retractions;
• enter publication agreements; and
• be accountable for ethical and scholarly conduct.

Responsibility for AI-assisted material remains with the human authors.

Authors must verify:

• factual accuracy;
• scientific accuracy;
• clinical accuracy;
• statistical accuracy;
• reference accuracy;
• data provenance;
• methodological validity;
• originality;
• appropriate attribution;
• ethical compliance; and
• consistency between the manuscript and underlying research.

Authors may not defend an error by stating that it was produced by an AI system. Use of an automated tool does not transfer accountability away from the authors.

Disclosure is particularly appropriate where AI was used for:

• generation of substantive manuscript text;
• translation affecting scientific meaning;
• data analysis;
• statistical modelling;
• code generation used in analysis;
• image generation or substantive image modification;
• classification or prediction;
• literature screening;
• data extraction;
• automated evidence synthesis;
• generation of research materials; or
• another contribution capable of materially affecting the scholarly work.

Disclosure should identify, where relevant:

• the tool or system used;
• the purpose for which it was used;
• the stage of the research or manuscript process;
• the nature of human verification; and
• any limitations relevant to interpretation or reproducibility.

Examples may include:

• spelling correction;
• grammar correction;
• basic formatting;
• minor readability improvement;
• reference-format conversion; or
• other low-level assistance comparable to ordinary editing software.

However, authors remain responsible for ensuring that such assistance does not alter scientific meaning or introduce errors.

If a tool substantially rewrites interpretation, generates claims, changes statistical meaning, or contributes materially to scholarly reasoning, that use should be disclosed.

Generative systems may produce:

• non-existent articles;
• incorrect author names;
• incorrect titles;
• incorrect journal names;
• fabricated DOIs;
• incorrect publication years;
• misrepresented findings; or
• references that do not support the cited claim.

Authors must verify references against reliable scholarly sources before submission.

Fabricated references may constitute a serious publication-integrity concern, particularly where they are knowingly retained or used to create a false evidence base.

Prohibited uses include fabrication of:

• research participants;
• patient records;
• interviews;
• survey responses;
• laboratory measurements;
• clinical observations;
• statistical outputs presented as actual analyses;
• ethics approvals;
• consent records;
• trial registrations;
• citations;
• reviewer identities;
• institutional affiliations; or
• other evidence presented as genuine research records.

Synthetic data may be scientifically legitimate in certain research designs, simulations, benchmarking studies, or methodological work, but it must be clearly identified as synthetic and described transparently in the methods.

Authors using synthetic data should describe:

• why synthetic data were used;
• how the data were generated;
• the software or model used;
• key generation parameters;
• whether empirical data informed the synthetic dataset;
• how the synthetic data were validated; and
• limitations affecting interpretation.

Synthetic observations must never be described as real patients, participants, biological samples, or clinical events.

Particular caution applies to:

• radiographs;
• MRI or CT images;
• angiography;
• histology;
• microscopy;
• clinical photographs;
• pathology images;
• gel images;
• medical-device outputs; and
• other images presented as research evidence.

AI must not be used to invent lesions, remove inconvenient findings, generate non-existent anatomical structures, or otherwise create a misleading representation of source evidence.

Scientifically legitimate image processing or reconstruction should be described sufficiently to permit editorial assessment.

Where an AI-generated image is used only as:

• a conceptual illustration;
• a graphical abstract element;
• an educational schematic;
• a non-evidentiary visualisation; or
• another clearly illustrative element,

it should be identified appropriately and must not mislead readers into believing it represents actual experimental, clinical, or patient-derived data.

Copyright, provenance, and third-party rights must also be considered.

Authors should verify:

• code syntax;
• statistical assumptions;
• variable definitions;
• data transformations;
• missing-data handling;
• model specification;
• output interpretation;
• reproducibility;
• software dependencies; and
• consistency with the stated methodology.

AI-generated code should not be used uncritically merely because it executes without an error.

Where AI materially contributes to computational analysis, authors should disclose that contribution and provide sufficient methodological detail for evaluation.

Relevant information may include:

• model or system name;
• version where available;
• access date where system behaviour may change over time;
• model configuration;
• input and prompting procedures;
• training or fine-tuning information where known and relevant;
• data sources;
• validation methods;
• performance metrics;
• human oversight;
• bias and fairness assessment;
• privacy safeguards;
• reproducibility limitations; and
• other information necessary to evaluate the research.

Depending on the study, authors may need to describe:

• system prompts;
• user prompts;
• prompt templates;
• few-shot examples;
• temperature or sampling settings;
• number of repetitions;
• response-selection procedures;
• human adjudication;
• output coding; and
• measures taken to assess variability or reproducibility.

Proprietary limitations should be disclosed rather than presenting an opaque workflow as fully reproducible.

Particular caution applies to:

• patient-identifiable information;
• protected health information;
• genomic data;
• unpublished research data;
• confidential institutional information;
• proprietary datasets;
• peer-review material;
• embargoed research;
• commercially confidential information; and
• other restricted material.

De-identification should be appropriate to the data involved; simply removing a participant's name may not sufficiently protect sensitive clinical or genomic information.

Reviewers must not upload confidential manuscript material to a generative-AI system where:

• the service may retain the content;
• the content may be used for training;
• confidentiality cannot be assured;
• third parties may access the information;
• intellectual property may be exposed; or
• the journal has not authorised such processing.

This includes:

• manuscript text;
• abstracts;
• tables;
• figures;
• supplementary data;
• unpublished results;
• author information; and
• reviewer correspondence.

Reviewers must not delegate the intellectual responsibility for peer review to an AI system.

AI tools may potentially support limited administrative or editorial functions where confidentiality and data governance are adequate, but they must not independently determine:

• whether a manuscript is scientifically valid;
• whether an author committed misconduct;
• which reviewer should prevail in a disagreement;
• whether a manuscript should be accepted;
• whether a manuscript should be rejected;
• whether an article should be retracted; or
• another final decision requiring accountable editorial judgement.

Final responsibility remains with the authorised human editor.

Detection systems may produce:

• false positives;
• false negatives;
• inconsistent results between tools;
• different scores after minor editing;
• language-related bias; or
• results that cannot establish who generated a particular passage.

Therefore, an automated score should not be treated as proof of misconduct without contextual assessment and appropriate supporting evidence.

Authors should ordinarily have an opportunity to explain material concerns before adverse integrity conclusions are reached.

AI-assisted text may:

• reproduce phrases from source material;
• closely paraphrase existing publications;
• generate unattributed summaries;
• combine ideas without clear source provenance; or
• create misleadingly original-looking text based on existing work.

Authors must check AI-assisted output for plagiarism and appropriate attribution before submission.

Authors should verify the rights status of:

• AI-generated images;
• AI-modified figures;
• generated code;
• translated material;
• adapted diagrams;
• training-derived content; and
• other generated material proposed for publication.

Serious concerns may include:

• fabricated data;
• fabricated participants;
• fabricated references;
• fabricated quotations;
• fabricated ethics documentation;
• fabricated peer reviewers;
• deceptive image generation;
• undisclosed synthetic evidence presented as empirical data;
• material plagiarism;
• confidentiality breaches;
• AI-generated peer reviews submitted without meaningful reviewer assessment;
• AI-generated author identities or affiliations; or
• deliberate concealment of material AI use that affects evaluation of the research.

Such concerns may be assessed under the journal's publication-ethics procedures.

Possible actions may include:

• request for clarification;
• addition or correction of an AI-use disclosure;
• correction of references;
• correction of figures or text;
• integrity investigation;
• Expression of Concern;
• retraction where the work is materially unreliable; or
• another proportionate scholarly-record action.

Failure to disclose minor language assistance does not automatically justify retraction. The journal considers the nature and consequences of the omission.

A suitable disclosure may follow this structure:

The authors used [tool/system, version if known] for [specific purpose]. All generated or AI-assisted content was reviewed and verified by the authors, who take full responsibility for the final manuscript.

Where AI was used as part of the research methodology, the Methods section should contain sufficient additional detail for scientific assessment and reproducibility.

• ICMJE Recommendations
• COPE Guidance and Resources
• Principles of Transparency and Best Practice in Scholarly Publishing

Because AI technologies and publication standards continue to evolve, JPMHR may revise this policy when substantive new ethical or methodological guidance emerges.

Reference to an external organisation does not imply membership, accreditation, certification, or endorsement unless independently established.