JAMA Network AI Policy August 2026: Four Prohibitions Every Clinical Author Must Know

In August 2026, the JAMA Network updated the AI guidance embedded in its Instructions for Authors across all affiliated journals. The update is not a full reversal of earlier policy, which already prohibited listing AI tools as authors and required disclosure of most substantive AI use. What is new is the explicit language closing off four specific categories of AI involvement entirely, with no disclosure pathway, no exceptions for minor use, and no room for interpretation. If you are submitting to JAMA, JAMA Internal Medicine, JAMA Cardiology, JAMA Oncology, or any of the other journals in the network, these four prohibitions now define what the editorial office expects before it even looks at your science.

It is worth pausing on what "across all affiliated journals" actually means here. The JAMA Network includes at least 13 titles that span most of clinical medicine: the flagship JAMA, along with JAMA Internal Medicine, JAMA Cardiology, JAMA Oncology, JAMA Neurology, JAMA Psychiatry, JAMA Surgery, JAMA Pediatrics, JAMA Dermatology, JAMA Ophthalmology, JAMA Otolaryngology-Head and Neck Surgery, JAMA Network Open, and JAMA Health Forum. A policy change embedded in the Instructions for Authors of that family touches a very large share of clinical authorship. Researchers who submit across specialties are now operating under the same set of rules regardless of which JAMA journal they target.

Why JAMA Moved to Explicit Prohibitions

Earlier JAMA guidance, like that of most major publishers, dealt with AI through a disclosure-first framework: tell us what you used, tell us what it touched, and take responsibility for the output. That framework worked reasonably well for manuscript text because the underlying principle is straightforward. If an author used a chatbot to tidy up a discussion section and discloses it clearly, the disclosure itself signals appropriate transparency. The human authors still conducted the study. They still wrote the draft. They are still accountable.

The problem is that disclosure does not work as a safeguard for every kind of AI use. For peer review, the concern is not transparency but confidentiality: a reviewer who uploads a manuscript to an external AI tool has already committed the breach before they ever write a word of disclosure. For clinical images, the concern is data integrity: a single AI-fabricated pathology slide or radiology image cannot be made acceptable by labeling it AI-generated, because the image itself would constitute fabricated primary data. For citations, the concern is accuracy: an AI-generated reference that does not exist cannot be disclosed into existence. For opinion writing, the concern is intellectual honesty: a letter or editorial credited to a named author should represent that person's own reasoning, not a model's output.

In each of these four cases, the harm is upstream of disclosure. That is why JAMA chose prohibition rather than a more elaborate disclosure regime. The practical consequence for authors is that these four areas now require a different kind of pre-submission check: not "did I disclose this?" but "did this happen at all?"

Prohibition 1: AI Has No Role in Peer Review

Peer reviewers invited by JAMA Network journals are now explicitly prohibited from entering manuscript text, figures, or any identifying content into an external AI tool or chatbot. This prohibition covers both the manuscript under review and the draft review itself. A reviewer who feeds the submission to a large language model to get a summary, or who uses one to draft or refine their comments, has violated the policy regardless of whether they disclose it to the editor.

The reason JAMA cites is confidentiality, and that reasoning deserves to be read carefully. When a manuscript is sent for peer review, it arrives as privileged material. The authors have not yet published it. They have not authorized its sharing beyond the editorial office and invited reviewers. The moment a reviewer uploads it, or any meaningful excerpt, to a third-party AI service, that material has left the controlled editorial environment. Depending on the service's data handling practices, it may be stored, logged, used for training, or retrievable by other users. That is a confidentiality breach even if the resulting review is excellent and fully human-edited before submission.

What does this mean for you as an author? It does not change what you submit. It changes what protections apply to your submission once it leaves your hands. Authors submitting to JAMA Network journals can now point to an explicit editorial commitment that their unpublished work will not be entered into external AI systems at the review stage. That is a meaningful assurance for authors working in competitive subfields, for those conducting clinical trials with commercially sensitive protocols, and for anyone whose preliminary findings could be distorted if they escaped the editorial process prematurely.

There is also an indirect practical implication. If you are invited to review for any JAMA Network journal, the prohibition is binding on you as a reviewer, not just on authors. Using AI to assist your review, even to summarize a long methods section, now violates the terms under which you agreed to review. The professional consequences of that violation could be significant.

Prohibition 2: Clinical Images, Illustrations, Video, and Audio

AI-generated or AI-manipulated clinical images, illustrations, video, and audio are now prohibited from JAMA Network publications unless the AI is the subject of the research itself and is fully disclosed. That narrow exception matters. A paper studying how AI performs in radiograph interpretation might legitimately include AI-generated outputs as research data, properly labeled, with full methodology. A clinical case report that includes an AI-generated representation of a patient's CT findings is a different matter entirely.

The concern here is obvious once stated plainly. Clinical images in medical journals are, in most contexts, representations of actual clinical findings. Pathology slides, echocardiogram frames, dermatologic photographs, ophthalmoscopy images, surgical field photographs: these are treated as primary data, not as illustrations in the decorative sense. If a reader sees a histopathology image in a JAMA paper and the image was synthesized rather than photographed from actual tissue, the scientific record has been falsified even if the surrounding text is accurate. The image claims to show something real. It does not.

This extends beyond obvious cases of fabrication. The prohibition also covers AI manipulation: running a real clinical image through a generative model to alter its appearance, enhance resolution, remove artifacts, or highlight features. Even well-intentioned modifications can change the medical meaning of an image. What looked like a benign finding might be altered in ways that make it appear pathological, or vice versa. The AMA has been conservative about image editing for years, and the August 2026 update extends that conservatism explicitly to AI-based manipulation.

The practical implication for clinical authors is that you need to be able to trace every clinical image in your submission back to its original capture. If a member of your team used any AI-based imaging enhancement, super-resolution, or generative fill tool on an image before inserting it into the manuscript, that image now fails the JAMA policy bar regardless of the reason the tool was used. This is worth confirming explicitly with co-authors who handle imaging, because the editing may have happened casually as part of a normal image workflow without anyone recording it as a compliance question.

Prohibition 3: AI-Generated Citations

The prohibition on AI-generated citations reflects a problem that medical editors have been documenting since early in the generative AI era. Large language models produce plausible-sounding references with real journal names, plausible author names, real-looking DOIs, and publication years that fit the topic. Many of them do not exist. Some are composites of real papers. Some have real titles and authors but wrong journals, wrong years, or fabricated page numbers. The model is generating statistically plausible text, not retrieving verified bibliographic records.

JAMA has now made explicit what was already implied by general research integrity standards: references cannot be generated by AI. Every citation in a submission must be traceable to a source the author has read or at minimum verified against a reliable database. This is not a new standard in spirit. Medical authors have always been expected to cite papers they can vouch for. The August 2026 update simply closes the loophole some authors were taking advantage of by treating AI citation suggestions as a starting point and then failing to verify them before submission.

The harder practical question is where literature-search AI tools fit. Tools like Elicit, Consensus, and Semantic Scholar surface real papers from real databases, not hallucinated titles. Using them to find candidate references and then reading and verifying each candidate before citing it should fall outside the prohibition, because the citation is ultimately attested by the author who read the paper. Where it becomes a problem is the pattern of importing suggested references without verifying that the paper exists, says what you think it says, and is correctly described in your manuscript. The prohibition is really targeting that second pattern, where the AI was the last step between the reference list and the submission.

Prohibition 4: Opinion Pieces, Letters, and Responses to Reviewers

The fourth prohibition covers a genre that receives less attention in AI policy discussions but carries its own form of integrity risk. Opinion pieces, editorials, letters to the editor, and responses to peer review are forms of academic discourse in which the named author is making a personal intellectual commitment. An editorial in JAMA signed by a named cardiologist is understood to represent that person's reasoned position on a clinical question. A letter to the editor challenging another author's methodology is understood to reflect the letter writer's own analysis.

JAMA's August 2026 policy prohibits using AI to draft or substantially generate any of these forms. The reasoning tracks the authorship principles the journal has maintained throughout: AI tools are not authors and cannot take accountability for published claims. When a letter or editorial is AI-generated, the signed author has effectively lent their name to content they did not produce, even if they reviewed and approved it. That is a different integrity issue from manuscript text, where AI-assisted drafting with appropriate disclosure is now normalized. In opinion writing, the named voice is the point. Substituting a model for that voice, regardless of disclosure, undermines the genre's function.

Responses to reviewers belong in the same category. When peer reviewers ask an author to explain their methodological choices, interpret a discrepancy, or justify their conclusions, the response is a direct intellectual exchange. Using AI to draft that response misrepresents the author's own understanding of the work. It also creates a practical risk: AI-generated reviewer responses tend to be formally correct but evasive, addressing the surface structure of the reviewer's comment without engaging with its specific concern. Editors and reviewers often notice this pattern, and it rarely improves the revision outcome.

The Disclosure Middle Ground: What Is Still Permitted

Beyond the four prohibitions, the JAMA Network policy continues to operate on a disclosure-first model for everything else. Grammar and spelling correction require no disclosure, consistent with the long-standing convention that copy-editing tools are part of normal manuscript preparation. But that narrow exemption for grammar tools is the only truly disclosure-free category. Most other AI-assisted manuscript work now requires a formal statement.

Translation is one area where AI use is permitted with disclosure. Many international clinical research teams write initial drafts in their primary language and then translate into English for submission. AI translation tools have become genuinely useful for this step, and JAMA's policy does not prohibit them. The disclosure requirement, however, applies to translation just as it does to language editing. Authors should note the tool used, confirm that a human reviewer with clinical knowledge reviewed the translated text for accuracy, and describe this in the acknowledgements section.

Research use of AI in data analysis, classification, natural language processing, or other analytical steps remains permitted under JAMA's policy when it is fully reported. This is the area where clinical researchers are most likely to be using AI in legitimate and scientifically appropriate ways: applying a language model to classify free-text patient reports, using an AI system to extract entities from electronic health records, or using a machine learning tool for image segmentation in a radiology study. These uses belong in the Methods section, described with the same specificity the journal would expect for any other analytical tool. The tool name, version, key parameters, and any validation steps should all appear there.

Data visualization is similarly permitted in most cases, but AI-generated figures should be clearly labeled and the generation process described. This is distinct from the clinical image prohibition, which targets images representing clinical findings. A flow diagram, schematic, or conceptual figure generated with AI assistance sits in a different category, closer to graphic design than to primary data representation. The key question is whether the figure claims to show real clinical data. If it does, the prohibition applies.

Where JAMA Wants Disclosure to Appear

JAMA's guidance on disclosure placement follows a two-category logic that matches the nature of the AI use rather than simply appending a generic statement at the end of every manuscript. Understanding this placement requirement matters because misplaced disclosure is almost as problematic as missing disclosure. An editor who finds AI acknowledgement buried in a section where it does not belong may doubt whether the authors understood what they were disclosing.

Research-related AI use belongs in the Methods section. This applies whenever AI touched the underlying research workflow: data collection, analysis, classification, search strategy, coding of qualitative data, or any other step that shaped the results being reported. The Methods section is where readers look to evaluate whether the methods were sound. If AI tools were involved in producing those results, they belong in the same section as the other methods. Hiding AI analysis in the acknowledgements misrepresents where in the research process it occurred.

Manuscript preparation AI use belongs in the Acknowledgements section. This covers language editing, translation, formatting assistance, and any AI tool used only to improve the manuscript text after the research was complete. The distinction between research AI and writing AI is not always perfectly clean, but the guiding question is whether the AI influenced the scientific content or only the presentation. Drafting assistance for the discussion is a gray area that may warrant a note in both sections.

JAMA submission systems may also include a dedicated AI disclosure field. Whatever you enter there should be consistent with what appears in the manuscript. Discrepancies between the submission form and the published text are a post-acceptance complication that takes time to resolve and can delay publication.

How JAMA's Policy Compares to NEJM, The Lancet, and BMJ

All four of the flagship general medical journals now have explicit AI policies, and they share most of their core positions. None of them lists AI tools as authors. All of them require disclosure of substantive AI use. All of them prohibit AI-generated citations. The differences show up in how each journal phrases its specific restrictions and how far it has moved toward explicit prohibition versus case-by-case disclosure.

The New England Journal of Medicine has maintained a disclosure-first approach for most AI use, but has been particularly direct about citations and peer review confidentiality, positions it shares with the updated JAMA guidance. The Lancet family, published by Elsevier, has broadly aligned with its publisher's group-level AI policy, which tilts toward prohibition for the same categories JAMA now restricts. BMJ has an active editorial conversation about AI in publishing, and its guidance has evolved more openly, with policy pages that acknowledge ongoing uncertainty about some edge cases.

The meaningful practical difference is that JAMA's August 2026 update made four prohibitions explicit and uniform across a 13-journal network in a single policy document. That degree of coordination is unusual. Most publisher-level policies still leave some room for individual journals to interpret requirements differently. If you are submitting across the JAMA family, you can now apply the same checklist to every journal in the group without having to parse journal-specific variations.

Authors who submit across multiple high-impact medical journals should be aware that policies at journals outside the JAMA Network may differ on the clinical image question in particular. Some journals outside the JAMA family have not yet made the image prohibition as explicit, particularly for schematic illustrations and conceptual diagrams that do not represent clinical findings. When submitting elsewhere, reading the current instructions rather than relying on JAMA's framework as a default is the only way to get this right.

A Pre-Submission Checklist for JAMA Network Submissions

Before you submit to any JAMA Network journal, the four prohibitions suggest a checklist that goes beyond the usual formatting review. This is not about formal compliance theatre. It is about having a clear answer ready if the editorial office asks.

The point of this checklist is not to create paperwork. It is to surface AI use that happened during manuscript preparation that a co-author might not have flagged because it felt routine at the time. The corresponding author is the one who signs the submission and carries the editorial accountability. That person needs to know what everyone on the team did, not just what was declared in the shared draft.

What This Means in Practice Going Forward

The August 2026 JAMA Network update represents the clearest signal yet that major clinical journals are moving past the disclosure-only framework for certain high-risk AI uses. The four prohibitions JAMA has now made explicit are almost certainly stable. Clinical image fabrication, citation hallucination, peer review confidentiality, and opinion ghostwriting represent genuine integrity risks that a disclosure statement cannot adequately mitigate. Other publishers are watching this policy evolution, and it is reasonable to expect similar explicit prohibitions to appear in the instructions of other flagship journals over the next year.

For most clinical authors, the behavioral change required by this policy is modest. Researchers who were already verifying citations, working with real clinical images, and writing their own opinion pieces are not being asked to stop doing anything. The policy formalizes what careful authorship already looked like. The change falls hardest on teams whose AI use was ad hoc, undocumented, or driven by the assumption that disclosure requirements were optional when no one was checking.

The most useful way to internalize these rules is not to memorize the four prohibitions as a list, but to think about the rationale behind each one. Peer review is confidential: do not export it. Clinical images represent real data: do not synthesize them. Citations must be verifiable: do not outsource them to a tool that invents. Opinion writing is personal: sign only what you wrote. Each rule follows from a principle that serious clinical authors already hold. The policy just makes the principle explicit and gives the editorial office a clear standard to enforce.

If you are preparing a submission to JAMA or any of its affiliated journals, the concrete next step is simple: read the current Instructions for Authors for your specific target journal rather than relying on remembered policy from a previous submission. The August 2026 update is embedded there, and the instructions for specialty journals within the network sometimes add requirements beyond the network-level floor.

Further Reading