Frontiers Built Its Own AI Tool for Reviewers, So They Would Not Need an External One

Frontiers’s policy follows a now-familiar shape — no AI authorship, disclose AI writing, check AI figures, keep manuscripts out of outside tools — with one genuinely distinctive move on the last point.

Authorship. “Authors should not list a generative AI technology as a co-author or author of any submitted manuscript”, because “Generative AI technologies cannot be held accountable for all aspects of a manuscript” — explicitly tied to not meeting ICMJE authorship criteria, the same four-part test BMJ applies directly.

Disclosure. Where AI generated the main text, that use “must be clearly disclosed in the acknowledgments”, and “this explanation must list the name, version, model, and source of the generative AI technology” — a more granular naming requirement than most publishers in this set ask for. Authors are encouraged, though not required, to upload their prompts and outputs as supplementary files.

Figures. “Figures produced by or edited using a generative AI technology must be checked” against the manuscript’s own data — a verification duty rather than an outright ban, which puts the responsibility on the author to confirm the figure still accurately represents what was actually found.

Peer review, and the distinctive part. Manuscript content must “not be uploaded to external generative AI tools”. Rather than stopping there, Frontiers provides reviewers an internal AI tool, in a closed environment, that summarises manuscripts and helps draft feedback — but “does not participate in editorial decision-making, and should not be used for that purpose”. Any use of an external tool needs prior discussion with the editorial team and, if approved, disclosure to authors in the review.

Building an internal tool rather than only banning external ones is a different strategy from every other publisher in this set: it accepts that reviewers already use AI assistance — Frontiers cites its own research finding a majority of reviewers do — and tries to channel that use into an environment where confidentiality can actually be controlled, rather than relying on a prohibition reviewers may or may not follow.

Separating the tool from editorial decision-making specifically closes the obvious next risk: a summarisation aid is one thing, and a tool that influences whether a paper is accepted is another, and the policy is explicit that the internal tool is confined to the first.

Requiring the model and source alongside the name and version, in the disclosure, is a finer level of detail than most publishers ask for, and it matters for the same reason naming a version matters elsewhere: two different products sharing a brand name can have very different failure patterns, and only a specific-enough disclosure lets that be checked later.

How it reaches a specific journal. Frontiers publishes the policy group-wide, covering its full portfolio from Frontiers in Immunology to Frontiers in Artificial Intelligence; individual sections or journals may layer their own editorial expectations on top without lowering the baseline described here.

For a reviewer specifically, the practical difference from most other publishers is access: Frontiers is, among the publishers in this set, the one that gives reviewers a sanctioned AI tool at all, rather than leaving them with a ban and no alternative.