The settled positions
ICMJE and COPE both hold that an AI tool cannot be an author. Authorship requires accountability for the work, including the ability to approve the final version and to respond to questions about its accuracy and integrity. A tool cannot do any of that.
Use must be disclosed. Most publishers now require a statement naming the tool, its version, and what it was used for. The usual place is the methods section or an acknowledgement, depending on whether the use touched the research or only the text.
The authors remain fully responsible for everything in the manuscript, including text a model produced. "The AI generated it" is not a defence against an error, a fabricated citation or plagiarised wording.
Where the real risks are
Fabricated references. Language models generate citations that are plausibly formatted and do not exist, or that exist but do not say what they are cited for. Every reference in an AI-assisted draft must be opened and checked against the source — not merely confirmed to resolve.
Confident error. Fluent prose signals nothing about accuracy. The failure mode is a sentence that reads authoritatively and is wrong in a way a non-specialist reader cannot detect.
Unattributed text. Model output can reproduce training material closely enough to constitute plagiarism, which remains the author’s responsibility.
Confidentiality. Unpublished manuscripts, patient data and commercially confidential material should not be entered into general-purpose tools, where the text leaves the author’s control.
Where it is genuinely useful
Language editing for non-native speakers, where the content is already fixed. Restructuring an argument the author has written. Drafting routine sections against supplied facts. Summarising the author’s own text. Checking a draft against a reporting checklist.
The common feature is that the author supplies the substance and verifies the output. Uses where the model supplies facts are the ones that produce retractions.
A practical routine
Keep a record of what was used and how, so disclosure is accurate rather than reconstructed. Verify every factual claim and every citation against a source. Read the final text as though a stranger wrote it — because in part one did.
In short
What to take from this.
- AI cannot be an author — ICMJE and COPE are explicit.
- Open and check every reference; fabricated citations are the characteristic failure.
- Never enter unpublished or patient data into a general-purpose tool.
Sources
- ICMJE Recommendations, on AI and authorship
- COPE position statement on authorship and AI tools
- WAME recommendations on chatbots and scholarly manuscripts
Related