Why an AI Tool Cannot Be an Author, and What Researchers Must Disclose

An AI system can assist with a manuscript, but it cannot be its author, and medical-publishing guidance keeps authorship human because authors must approve the final work, disclose conflicts, answer questions, correct errors, and accept public responsibility. A software output cannot perform those duties.

Key points#

• ICMJE's recommendations, updated in January 2026, say AI tools should not be listed or cited as authors.
• Human authors are accountable for accuracy, attribution, permissions, originality, and the final wording, including AI-assisted material.
• Disclosure should identify the tool and explain its purpose, with the location adapted to the journal's policy.
• Unpublished manuscripts and confidential peer-review files should not be uploaded to a system unless confidentiality is assured and the journal permits it.
• A disclosure is not a quality certificate. Authors must still verify every claim, citation, calculation, table, image, and analytical decision.

Authorship is a responsibility structure#

Authorship is sometimes treated as a reward for producing text. Journal standards define it more broadly. Under ICMJE criteria, an author makes a substantial intellectual contribution, helps draft or critically revise the work, approves the version to be published, and agrees to be accountable for all aspects of the work.

An AI tool cannot give legally meaningful final approval. It cannot declare a financial interest, consent to a license, preserve source data, answer an editor's query from memory, or lead an investigation when a figure is challenged. It cannot promise to correct the record. Because it cannot carry the obligations, it cannot receive the authorship role.

The human authors do not divide responsibility with the tool. They retain it. Naming a model in the author list would obscure that chain of accountability rather than clarify it.

Assistance can be allowed without authorship#

The prohibition on AI authorship is not a universal ban on AI assistance. Depending on the journal and institution, tools may be used for language editing, code suggestions, idea organization, translation, literature triage, image processing, or drafting, though each use carries different risks and may trigger different reporting requirements.

A spelling correction is not equivalent to generating a discussion section. A coding assistant that reformats a plotting script is not equivalent to a system that proposes the statistical model, and a tool used to create a graphical abstract raises different questions from one used to identify candidate papers.

The relevant questions are therefore functional:

• What system and version were used?
• On what date or during what stage of the project?
• What information was supplied to it?
• What output did it produce?
• Which human verified and revised that output?
• Did the use affect methods, results, interpretation, or only presentation?
• Can another researcher understand or reproduce the consequential part?

Journal policy controls where these details belong. ICMJE recommends disclosure in the cover letter and in the submitted work where applicable, with writing assistance acknowledged; methods may be the right place when a tool performed part of an analysis or data workflow. An acknowledgment may suit language assistance. A supplement or repository can hold prompts, model settings, or validation details when those are needed for reproducibility.

Four risks disclosure does not solve by itself#

Plausible but false statements#

Generative systems predict output rather than verify truth. They can state a wrong mechanism, invent a trial result, merge two studies, or supply a citation that does not exist. Fluent language can make these errors harder to notice.

Human verification means opening the source, checking that it supports the precise claim, and confirming bibliographic details. A link that resolves is not enough. The source has to say what your manuscript attributes to it.

Hidden borrowing and copyright#

An output may resemble protected text or omit the source of an idea. Authors remain responsible for correct quotation, paraphrase, attribution, and permissions. A statement that AI was used does not excuse plagiarism or unauthorized reproduction.

Confidentiality and data protection#

Submitted manuscripts are privileged communications. Peer reviewers and editors may receive unpublished methods, results, personal information, or commercially sensitive details. ICMJE says manuscripts should not be uploaded to AI systems where confidentiality cannot be assured without explicit author permission, and reviewers should seek journal permission before using AI to facilitate review.

The same caution applies before submission. Patient information, participant-level data, confidential protocols, grant applications, and proprietary code should not be sent to an external tool merely because the interface is convenient. Check the institutional agreements, the retention settings, the training-use terms, the access controls, and the applicable privacy rules before anything leaves your machine.

Bias presented as synthesis#

An AI-generated literature summary may omit contrary studies, overrepresent accessible English-language sources, or reproduce biases in its training material. Authors must design and document searches appropriate to the research question. A chatbot conversation is not a systematic search and should not be presented as one.

A practical verification ledger#

For substantive AI use, a simple internal ledger can preserve accountability. It can record the tool, version, date, purpose, input class, output destination, reviewer, verification method, and disposition. Sensitive prompts do not need to be copied into an insecure file, but the workflow should be reconstructable.

Verification should match the task:

• Text claims require comparison with primary or authoritative sources.
• Citations require title, author, journal, year, identifier, and claim-level checks.
• Code requires review, tests, known-answer examples, and version control.
• Statistical output requires an independent calculation or validated pipeline.
• Images require provenance, manipulation disclosure, and checks that no unsupported feature was added or removed.
• Translations require review by someone competent in both the source language and the scientific context.

Defend the final artifact without pointing at the system as the authority.

Peer review has a higher confidentiality bar#

Reviewers are entrusted with work that does not belong to them and may not yet be public. Uploading it can disclose the manuscript to a third party, create retention outside the journal's systems, or permit reuse under terms the authors never accepted.

Before any AI-assisted review, read the journal policy, request permission where it is required, and confirm that the tool's confidentiality protections are adequate. If you cannot satisfy those conditions, do not upload the manuscript. Reviewers also remain responsible for the accuracy, fairness, originality, and tone of the review and should disclose permitted use to the editor.

Editors and publishers face the same problem at scale. Automated screening or summarization should have a documented purpose, human oversight, access control, and a policy visible to authors. The convenience of processing does not cancel the duty to protect a confidential submission.

Policy differences are real#

ICMJE and the World Association of Medical Editors provide influential principles, but individual journals can be stricter. Some permit limited language assistance with disclosure. Some prohibit generated images. Some require methods-level detail for analytical use. Some restrict AI use by reviewers entirely.

Rules also change quickly. Before you submit, check the target journal's current instructions, the publisher's policy, institutional research rules, funder requirements, and any study-specific data agreements, because a disclosure drafted for one journal should not be assumed sufficient for another.

A defensible disclosure pattern#

A useful statement is concrete and proportionate. It might identify the named tool and version, state that it was used to improve grammar in specified sections, and explain that you reviewed and revised all output and take responsibility for the final text. If the tool assisted with code or analysis, the statement should describe the task, validation, and where reproducibility materials can be found.

Vague language such as “AI was used in preparing this work” is hard to evaluate. Excessive detail about trivial autocorrection can also obscure the important use. The goal is to let an editor and reader understand what intellectual or technical work the system affected.