JAMA publishes updated guidance on the use of AI assistance - here.
See a Linked In response - here.
Here are two, 400-word responses, first by Chat GPT and then by Claude Opus.
##
Response to the JAMA Editors: AI, Authorship, and Responsibility
The new JAMA Network guidance on artificial intelligence is thoughtful in its emphasis on accuracy, transparency, and author accountability, but it creates a tension that deserves reconsideration. The guidance begins from a strong principle: authors remain fully responsible for everything submitted, including AI-assisted content, and must verify its accuracy and integrity. AI may therefore be used for research, literature searches, manuscript preparation, translation, and data visualization when appropriately disclosed. Yet JAMA draws a categorical line around Opinions, Letters, Comments, and responses to reviewers: AI may not be used to draft them because “human insights are prioritized.”
That distinction is understandable but increasingly difficult to defend.
The editors identify genuine problems. JAMA has encountered fabricated references and a marked increase in formulaic, low-value submissions, sometimes from authors lacking subject-matter expertise. These are important editorial concerns. But they are fundamentally problems of quality, accuracy, expertise, and accountability, rather than problems caused simply by the presence of an AI tool.
Stefano Natangelo has identified the central inconsistency particularly well: if authors truly bear responsibility for every claim, citation, and argument, it is unclear why responsibility becomes insufficient only for selected manuscript categories. A superficial Opinion remains superficial when composed entirely by hand; conversely, an experienced investigator may use an LLM to organize, challenge, compress, or improve an argument that remains unmistakably the investigator’s own. As Natangelo observes, what ultimately matters is whether the references exist, the sources support the claims, and the argument withstands scrutiny.
There is also a practical problem: the boundary between “drafting,” “revision,” and permissible assistance will become progressively less meaningful as AI is embedded in word processors, search engines, reference systems, and ordinary writing software.
JAMA should therefore consider replacing categorical authorship bans with a simpler standard: disclose substantive AI assistance, require authors to verify every factual assertion and citation, and hold them completely accountable for the final work. Editors should reject formulaic, derivative, inaccurate, or intellectually empty submissions—regardless of whether they were produced by a human alone or with AI.
The enduring editorial question should be not Who typed the sentences? but Are they accurate, original, useful, and worth publishing?
##
##
Now by Claude OPUS.
##
.
JAMA Tightens the Screws on AI: What It Says, and What It Can't Do
On August 10, Flanagin, Perlis, and Bibbins-Domingo updated the JAMA Network's author guidance. The architecture is unchanged — authors own every word they submit — but a layer of prohibitions now sits on top. AI may no longer generate or format references; use a reference manager, and check them. AI may not draft Opinion manuscripts, Letters to the Editor, online Comments, or responses to reviewers, because "human insights are prioritized." AI-created or -manipulated clinical images, illustrations, video, and audio are barred unless they are the object of study. Research use, literature searching, manuscript preparation, translation, and data visualization remain permitted with disclosure: platform, version, manufacturer, dates of use, which portions, and an attestation of responsibility. Grammar and spell-checking require no disclosure at all.
The reference rule is the strongest part. The editors report receiving manuscripts with nonexistent citations, and cite an audit of 2.5 million papers. That is a documented failure mode with a clean remedy.
The rest is harder to defend. The editorial's opening principle — authors are accountable for all content, however produced — should make type-by-type bans unnecessary. If a signed attestation covers the references, the sources, and the argument, what does a ban catch that reading the manuscript does not? A vacuous Opinion typed by hand is still vacuous. The asymmetry gives the game away: AI is permitted where actual patient data are at stake, and forbidden where the stakes are rhetorical.
There is also the enforcement problem nobody says out loud. How would they know? A competently prompted draft is not detectable, and detectors are unreliable enough that acting on them is its own integrity risk. So the binding constraint is author candor — and here the editorial cites, at reference 18, the finding that authors already conceal AI use because they fear biased editorial assessment. The response is to add a category of use that is now punishable. That predictably drives disclosure down, not up, and it does so in exactly the categories where an honest note would have cost nothing.
The permitted/prohibited line is also dissolving under the editors' feet. Grammar checking is exempt and needs no disclosure; translation is permitted but must be disclosed. Both are now the same transformer, often in the same word processor.
The real problem is volume — a flood of formulaic, low-value submissions. That is a quality problem, caught by reading. Say so plainly, and reject on the merits.
