Generative Artificial Intelligence Policy

Policies

Generative Artificial Intelligence Policy

 

Rules for authors, reviewers, and editors on the use of AI tools in manuscript preparation and evaluation.

Artificial intelligence is both a subject of research in this journal and a tool that may be used while preparing a manuscript. This policy addresses the second use. Requirements for reporting AI as the object of study are set out in the final section.

The policy follows the ICMJE Recommendations, the COPE position statement on authorship and AI tools, and the WAME Recommendations on Chatbots and Generative AI. It applies to large language models, chatbots, image generators, and any other generative system, whether commercial, open source, or embedded in another application.

The Governing Principle

An artificial intelligence tool cannot be an author. Authorship carries accountability for the accuracy, integrity, and originality of the work, and for responding to questions about it after publication. A generative system cannot hold or discharge that responsibility, cannot approve the final version, and cannot declare competing interests.

It follows that the human authors remain fully responsible for every word, figure, number, and citation in the manuscript, including any produced with AI assistance. "The tool generated it" is never a defence against a finding of plagiarism, fabrication, or error.

1. Permitted Uses — Disclosure Required

The journal recognises that generative tools can reduce the disadvantage faced by researchers writing in a second language. The following uses are acceptable provided they are disclosed:

Language and Readability
Grammar correction, spelling, punctuation, sentence restructuring, and improvement of clarity in text the authors wrote themselves.
Translation
Translating the authors' own draft into English, provided the authors verify that the scientific meaning is preserved.
Reference Formatting
Converting a verified reference list into the required citation style. Every reference must still be checked against the original source.
Analysis Code Assistance
Help with writing or debugging statistical or image-processing code, where the authors have validated that the code performs as described in the Methods.
Literature Search Support
Identifying candidate references, provided each is retrieved, read, and verified in the original database before citation.
Schematic Illustration
Producing a diagram or graphical abstract that carries no research data, labelled in the legend as AI-generated.
2. Prohibited Uses
These uses are not permitted under any circumstances
Generating or altering research images. No clinical or experimental image may be created, denoised, super-resolved, in-painted, reconstructed, or otherwise modified by a generative model unless that processing is itself the subject of the study and is fully described in the Methods.
Presenting synthetic images as authentic. Images produced by generative models, including GAN or diffusion-based synthesis, must never be presented as real patient data.
Fabricating data, results, or statistics. Including the generation of plausible-looking values, tables, or patient records.
Producing references. Generative tools routinely invent citations that do not exist or attach real authors to papers they never wrote. Any fabricated reference found in a submission is treated as a serious integrity failure.
Writing substantive scientific content. The interpretation, discussion, conclusions, and clinical recommendations must be the authors' own intellectual work.
Uploading identifiable patient data. Submitting patient information, clinical images, or DICOM files to a public AI service is a breach of patient confidentiality and may violate data protection law.
Listing an AI tool as an author or co-author, or in the Acknowledgements as though it were a contributor.

Manuscripts that appear to have been generated substantially by an AI system, and manuscripts originating from paper mills using such systems, are rejected and reported to the authors' institutions.

3. How to Disclose

Any use of a generative tool beyond basic spelling and grammar checking built into word processing software must be declared in a dedicated section headed Declaration of Generative AI Use, placed after the Conclusion and before the Acknowledgements. The declaration must state the tool and version, the date or period of use, the specific purpose, and confirmation that the authors reviewed and take responsibility for the output.

Model declaration
"During the preparation of this manuscript the authors used [TOOL NAME, VERSION] in [MONTH YEAR] in order to improve the grammar and readability of text originally drafted by the authors. After using this tool the authors reviewed and edited the content as required and take full responsibility for the content of the published article."
Where no tool was used
"The authors declare that no generative artificial intelligence tool was used in the preparation of this manuscript."
4. Obligations of Reviewers and Editors
Manuscripts Must Not Be Uploaded
A manuscript under review is a confidential document belonging to its authors. Uploading any part of it — text, tables, figures, or supplementary files — to a generative AI service transfers that content to a third party and constitutes a breach of confidentiality, irrespective of the provider's stated data policy.
Review Reports Must Be the Reviewer's Own
The critical assessment of a manuscript is the intellectual contribution the journal is asking for. A report generated by an AI system is not a peer review. Limited use of a tool to improve the language of a report the reviewer has written must be disclosed to the editor.
Editorial Decisions Are Human Decisions
No editorial decision at this journal is made or delegated to an automated system. Editors may use screening tools for similarity and image integrity checks, but the judgement and the accountability remain with the editor.
Detection Tools Used With Caution
AI-detection software is unreliable and biased against non-native English writing. The journal does not reject a manuscript on the output of a detector alone; a concern is raised with the authors and assessed on the evidence as a whole.
5. Reporting Artificial Intelligence as the Subject of Research

Studies that develop, validate, or evaluate an AI or machine learning system for imaging are firmly within the journal's scope and are held to established reporting standards. Submissions must follow CLAIM (Checklist for Artificial Intelligence in Medical Imaging) and, where a prediction model is involved, TRIPOD+AI. Trials of AI interventions must follow CONSORT-AI or SPIRIT-AI; diagnostic accuracy studies must follow STARD-AI.

Data provenance. Source of images, acquisition parameters, scanner models and vendors, inclusion and exclusion criteria, and the ethics approval covering dataset use.
Data partitioning. Size and composition of training, validation, and test sets, with explicit confirmation that no patient appears in more than one partition.
Reference standard. How ground truth was established, the qualifications of the annotators, and inter-observer agreement.
Model description. Architecture, pre-training, hyperparameters, augmentation, hardware, and software versions in sufficient detail for replication.
Performance reporting. Metrics with confidence intervals, calibration where relevant, and comparison against an appropriate human or clinical baseline.
External validation. Results on data from an institution or scanner not used in training, or an explicit statement that external validation was not performed.
Bias and subgroup analysis. Performance across relevant subgroups, and discussion of dataset shift and generalisability.
Availability. A statement on access to code, trained weights, and data, and any restriction that applies.
Synthetic data. If generated images were used for training or augmentation, the method and proportion must be stated and the synthetic images clearly labelled.
Limitations and clinical claims. Claims of readiness for clinical deployment must be proportionate to the evidence presented.

Commercial AI products evaluated in a submission must be named, with version number, and any relationship between the authors and the vendor declared under the journal's competing interests policy.

6. Consequences of Non-Compliance
Before publication. Undisclosed use identified during review results in the manuscript being returned for a full declaration, or rejected where the use falls in the prohibited category.
After publication. The case is handled under the Correction and Retraction Policy. Fabricated data, references, or images lead to retraction and notification of the authors' institution.
Reviewers and editors. Breach of confidentiality through an AI service ends the individual's involvement with the journal and is reported to their institution where warranted.
A Policy Under Review

Generative technology is changing faster than publishing norms. This policy is reviewed at least annually by the Editorial Board and updated as guidance from COPE, ICMJE, and WAME develops. Authors should consult the version current at the date of their submission. Questions may be directed to [EDITORIAL EMAIL].

Radiology Research Journal (RRJ) · Generative AI Policy · Version 1.0, [MONTH YEAR]