Medical editors warn a chatbot's citations may not be real

The writer's problem

A medical-journal author who used a chatbot to help draft a literature summary faces a narrower version of a familiar problem: not just whether the tool can be an author, but whether the citations it offered actually say what it claims they say. The World Association of Medical Editors addresses both questions in its recommendations on chatbots and generative AI in scholarly manuscripts, revised from an original January 2023 version.

What the documents show

WAME's first recommendation states that chatbots cannot be authors because they cannot give final approval of a manuscript or be held accountable for its accuracy, and that they have no independent legal standing to sign a conflict-of-interest statement. Its second recommendation requires authors to disclose chatbot use and, where a chatbot generated new text, to specify the prompts used. Its third recommendation is the sharper point for a working writer: authors are responsible for a chatbot's accuracy and for ensuring that the cited sources support the chatbot's statements, an explicit acknowledgment that a chatbot can attach a citation to a claim the citation does not actually back. A fourth recommendation extends disclosure to editors and reviewers who use chatbots in their own communications.

The editorial choice

Read together, these recommendations push an editor toward requiring not just a disclosure statement but a citation-by-citation check before publication, since WAME frames unverified chatbot citations as a plagiarism and misconduct risk rather than a minor error class. That is a heavier verification step than a simple disclosure line, and it is WAME's own recommended practice, not a universal journal requirement.

What stays with the author

The document is unambiguous that responsibility for a chatbot's inaccuracies, including fabricated-sounding but false citations, sits with the human author who submitted the work, not with the tool or its developer. Nothing on WAME's own site gives authors a technical method for catching a fabricated citation beyond manual verification, which remains the author's own labor.

• Has every citation a chatbot supplied been opened and checked against the claim it is attached to?
• Have the prompts used to generate any new text been recorded, as WAME's second recommendation asks?
• Would an editor's current screening process catch a fluent citation that does not actually exist?

WAME's contribution to the authorship debate is less about credit than about a specific failure mode: fluent, well-formatted citations that a human still has to verify one by one.