Artificial intelligence is now part of medical writing whether journals like it or not. The relevant question is no longer whether doctors use these tools, but which uses are legitimate, which must be disclosed, and which will get a paper retracted.
The good news is that the major bodies broadly agree with each other. This article sets out where the line sits and why.
The governing principle: AI is a tool, not an author. You remain fully responsible for every claim, every number and every reference in your manuscript, regardless of what produced the first draft. If you cannot defend a sentence, it should not be in your paper.
What the Major Bodies Say
Four sources set the expectations most journals follow.
The International Committee of Medical Journal Editors holds that AI tools cannot meet authorship criteria, because authorship requires accountability, approval of the final version and the ability to answer questions about the work. Authors must disclose AI use and remain responsible for the accuracy of everything submitted, including checking that AI generated text is free of error and plagiarism.
The Committee on Publication Ethics takes the same position on authorship and emphasises that AI tools cannot take responsibility for the submitted work, so their use must be declared transparently.
The World Association of Medical Editors recommends disclosure of the tool used and how it was used, and stresses that authors are responsible for verifying facts and references.
Major publishers including Elsevier, Springer Nature, Wiley, Taylor and Francis and the JAMA Network have all issued policies aligned with these positions, though they differ in the detail of where the disclosure must appear.
Policies are moving quickly. Every statement in this article should be checked against the current instructions to authors of your specific target journal before you submit. What was acceptable eighteen months ago may now require an explicit statement in a named section.
Uses That Are Generally Accepted
These are broadly permitted, usually with disclosure:
- Language editing and grammar correction of text you wrote
- Improving readability of a section that is technically correct but badly expressed
- Restructuring your own draft for flow, without adding new claims
- Translating your own text into English
- Reference formatting into a journal specified style
- Explaining a statistical concept to you, so that you understand what your analysis is doing
- Generating a plain language summary from your own abstract
- Screening support in systematic reviews, where this is described in the methods and human verification is documented
The common thread is that the intellectual content originates with you and the tool operates on it.
Uses That Are Not Acceptable
- Generating data, results or statistics. This is fabrication, one of the most serious forms of research misconduct, and it ends careers.
- Generating or altering images. AI generated figures presented as real clinical or experimental images constitute falsification. Most publishers prohibit AI generated images outright except where the image itself is the subject of the research.
- Producing citations without verification. Language models invent plausible references complete with volume numbers and page ranges. Submitting them is at best negligence.
- Listing an AI tool as an author or in the author contribution statement as though it were a person.
- Writing an entire manuscript from a prompt and submitting it as your own scholarly work.
- Uploading identifiable patient data to a public AI service.
- Running a manuscript you are peer reviewing through an AI tool. That manuscript is confidential and is not yours to disclose to a third party service.
The Patient Data Problem
This is the risk Indian doctors most often overlook, and it is a legal exposure rather than merely an etiquette question.
When you paste content into a consumer AI service, that content leaves your institution. Depending on the product and settings, it may be retained on servers outside India, reviewed by staff, or used to improve the model. Your ethics committee approval almost certainly did not contemplate this. Your institution data policy probably prohibits it. India's Digital Personal Data Protection Act treats health data as personal data with obligations attached.
Practical rule: never paste names, hospital registration numbers, contact details, full dates, raw clinical images or unredacted records into any AI tool. Use fully de-identified aggregate data if you use these tools for analysis at all, and check what your institution permits before you do.
The same applies to unpublished manuscripts you have been sent to review, to grant applications you are assessing, and to another investigator unpublished data shared with you in confidence.
Hallucinated References Are the Most Common Way Doctors Get Caught
Language models produce fluent text that is optimised to look correct rather than to be correct. Their most dangerous output in a medical context is a citation that does not exist: a real journal, plausible authors, a title that fits your argument perfectly, and no such paper.
Related failure modes include real papers cited for claims they never made, correct references attached to the wrong statement, and outdated guidance presented as current.
The defence is straightforward and non negotiable. Open every reference. Confirm the DOI or PubMed identifier resolves. Confirm the paper actually says what you have attributed to it. If you did not read it, do not cite it.
Where AI Genuinely Helps a Busy Clinician
Used within these limits, AI is a real productivity gain:
- Turning your rough notes into readable prose that you then correct
- Tightening a discussion section that has grown bloated
- Explaining why a reviewer statistical objection matters, so you can answer it properly
- Drafting a cover letter from your abstract
- Checking whether your manuscript covers each item on a STROBE or CONSORT checklist
- Producing a first draft of a response to reviewers, which you then verify against the actual comments
None of this replaces the thinking. It reduces the friction between your thinking and the page.
A Workflow That Keeps You Safe
- Do the research yourself. Question, protocol, ethics approval, data collection and analysis are yours.
- Draft the scientific content yourself, even roughly. The argument must be your argument.
- De-identify anything before it touches a tool. No exceptions.
- Use AI for language and structure, not for facts.
- Verify every reference at source. Every one.
- Read the final manuscript line by line and satisfy yourself that you could defend each sentence in a viva.
- Record what you used and how, so that writing the disclosure later is trivial.
- Check your target journal current policy and write the disclosure it asks for.
Our companion article on how to write an AI use declaration gives sample wording for each of these scenarios.
What Happens If You Get It Wrong
Consequences scale with the seriousness of the breach. Undisclosed language editing is likely to result in a request to add a disclosure. Fabricated references usually mean rejection, and if discovered after publication, a correction or retraction. Fabricated data or images can mean retraction, notification of your institution, and referral to research integrity processes. A retraction is permanent and public, and it is attached to your name for the rest of your career.
Set against that, the cost of doing this properly is an hour of reference checking and three sentences of disclosure.
Frequently Asked Questions
Can I list ChatGPT as a co author on my paper?
No. The International Committee of Medical Journal Editors, the Committee on Publication Ethics and every major medical publisher agree that AI tools cannot be authors, because authorship requires accountability for the work and the ability to approve the final version and respond to questions about it. A software tool cannot take responsibility. Its use is disclosed instead, typically in the methods or acknowledgements.
Is it cheating to use AI to improve my English?
No. Language editing has always been permitted, whether done by a human editor or a tool, provided the scientific content is yours. Most publishers treat AI assisted language polishing as acceptable, and many still ask you to disclose it. What is not permitted is having a tool generate the scientific content, the analysis or the interpretation and presenting it as your own reasoning.
Can I paste my patient data into an AI chatbot for analysis?
You should not paste identifiable patient data into any public or consumer AI service. Data entered may be retained, processed outside India and in some configurations used for model training. This risks breaching your ethics committee approval, your institution data policy and the Digital Personal Data Protection Act. If you use AI for analysis at all, use fully de-identified data and check whether your institution permits it.
Will a journal detect that I used AI?
AI detection tools are unreliable in both directions and are not a safe basis for either accusation or reassurance. The practical risks are different: fabricated references that do not exist, statistical claims that do not match your data, and text duplicated from other sources. These are found by editors, reviewers and plagiarism software, and they lead to rejection or retraction regardless of how the text was produced.
Not Sure Where the Line Is?
We help doctors develop manuscripts that meet current journal policies on authorship, disclosure and research integrity.
Talk to Us →This article is general educational guidance, not legal advice. Journal policies, regulatory rules and institutional requirements change. Always check the current instructions to authors of your target journal and the current guidance of your institution, the NMC and the ICMR before acting.