JAMA’s Updated AI Guidance: What It Means for Authors Using Evidence Synthesis Tools

On August 10, 2026, JAMA and the JAMA Network published updated guidance for author use of AI in medical publication, authored by Annette Flanagin et al. It’s the second major revision to JAMA’s AI policy since 2024, and it tightens restrictions in a few specific places worth understanding before you seek publication of your next manuscript.

The core message hasn’t changed since 2024: authors are responsible for the accuracy of everything they submit, AI-assisted or not. What’s new is where JAMA has drawn firmer boundaries, and why.

What Changed

JAMA’s guidance table separates AI use into three tiers: allowed with disclosure, allowed without disclosure, and not permitted. A few line items matter most for anyone doing literature reviews:

Literature search is acceptable, with disclosure and accuracy checking. AI-assisted search remains fine, as long as it’s described in the Methods section and every result is verified.

Bibliographic data and reference management is now discouraged outright. This is the headline change. JAMA is telling authors not to use AI, including LLMs, to generate or format references, and to use standard reference managers instead. The reasoning is simple: these tools “may cause problems with the accuracy and integrity of cited references.”

Manuscript preparation remains acceptable, with disclosure and an accuracy check. Drafting assistance is still considered acceptable. Opinion pieces, letters, and online comments are not, because JAMA wants human insight prioritized in those formats specifically.

Data formatting and graphic visualization of research data are acceptable, with disclosure in the figure legend. Charts, figures, and visualizations generated or assisted by AI are fine, as long as the tool and version are named where the figure appears.

Clinical images, illustrations, video, and audio are not permitted unless they’re part of a formally designed and disclosed research method. 

Peer review input is also off-limits, entering a manuscript or review summary into an external AI tool is treated as a confidentiality violation.

Every permitted use carries the same disclosure requirement: name of the tool, version, manufacturer, date of use, a description of what it was used for, and confirmation that the author takes responsibility for the accuracy of what it produced.

How Would You Implement within Nested Knowledge?

Most of what JAMA’s guidance addresses sits upstream of where a systematic review platform typically operates, however, Nested Knowledge has several capabilities which are subject to this guidance: literature search, reference handling, and manuscript drafting.

Literature Search: “acceptable, with disclosure”. Smart Search generates a starting Boolean string and helps explore PICO terminology and MeSH alignment, but it is not intended to finalize a search strategy unsupervised. Every string it drafts is meant to be reviewed, refined, and signed off by an information specialist or author before it is run.Per JAMA’s requirement: usage must be disclosed and documented, checked by a human author, and the author(s) are accountable for the final search strategy.

References: Nested Knowledge does not process references with AI. JAMA’s concern is AI generating or formatting citations from scratch, a generative task that invites hallucination. That isn’t what happens in a systematic review pipeline built on Nested Knowledge. Records in a Nest come from indexed sources, PubMed, Embase, ClinicalTrials.gov, DOAJ, EuropePMC, MAUDE, OpenAlex, or imported files, and citations are retrieved and formatted using strict, deterministic logic. Every record traces back to a real, retrievable source throughout the duration of the entire process.

Manuscript preparation and data visualization: acceptable, with disclosure. Smart Abstract and Synthesis Chat can draft structured content, abstracts, summaries, narrative sections, and Synthesis can generate dashboards, forest plots, and PRISMA diagrams. Under JAMA’s guidance, none of this is off-limits. All have the same requirement: name the tool you use  in the Acknowledgement section or figure legend, check for accuracy (human-in-the-loop, always), and the authorial ownership.

With Rapid Adoption, Comes Great Accountability

At its heart, JAMA’s update is not really about AI capability, but rather about accountability catching up to adoption. The editorial is explicit in saying that disclosure policy alone cannot stop fabricated content from reaching submission, which is why some uses moved from “disclose it” to “unacceptable.” For example, Cochrane’s RAISE framework applies similar standards, standards we hold Nested Knowledge’s own AI features to: disclosed, verifiable, and never presented as a substitute for human expertise.