AI Rules for Peer and Grant Reviewers: What’s Allowed

Most guidance on AI in research is aimed at authors. Reviewers get far less attention, even though the rules for them are stricter.

The reason is simple. A proposal or manuscript under review is confidential work shared in trust. Pasting it into the wrong tool can break that trust, whatever the reviewer’s intentions.

The Two Principles Behind the Rules

The European Research Council (ERC) set out its position in guidelines dated March 23, 2026, for panel members and remote reviewers. It rests on two principles:

• Non-delegation: the judgment about a proposal’s merit must stay with the human reviewer.
• Privacy and confidentiality: a proposal must not be sent to an external AI system.

Most reviewer rules from funders and publishers follow one or both.

What the ERC Allows and Does Not Allow

Even a locally run model cannot be used to summarize a proposal, because that breaks the non-delegation principle.

What NIH and Publishers Say

NIH prohibits reviewers from using generative AI to analyze and critique grant applications. Its notice says uploading content from an application to an online generative AI tool violates peer review confidentiality, because AI tools offer no guarantee of where data are sent, saved, viewed, or used later.

Publishers take a similar line for manuscripts. Springer Nature asks reviewers not to upload manuscripts into generative AI tools, and Elsevier asks reviewers not to upload their review reports into AI tools, even to improve language.

Notice that the rules differ. The ERC allows language polishing of your own review under conditions, while Elsevier’s guidance is stricter. That is why you should check the policy of each funder or journal every time.

A Practical Approach for Reviewers

1. Read the policy first. Check the funder’s or journal’s current guidance before using any AI tool.
2. Keep proposal content out. Never paste proposal or manuscript text, titles, or distinctive details into a tool.
3. Do the judging yourself. Use AI, if at all, only for language or general background research.
4. Check your own notes. If you polish language, make sure your text contains no confidential detail.
5. Ask when unclear. A short email to the editor or program officer settles most doubts.

Why Tool Terms Still Matter

Look again at the ERC’s condition for internal systems: no storage by the provider, no use for training, and no access by unauthorized persons. Those are the same questions worth asking of any tool you use for the allowed tasks, such as polishing your own review.

A tool that states its data terms clearly, and backs them up in writing, is easier to defend if anyone asks.

Where Trinka’s Confidential Data Plan Fits

Trinka’s Confidential Data Plan processes content in memory with zero data retention and does not use it to train any model. For the allowed tasks, such as checking grammar in a review you wrote yourself, those terms line up with the kind of conditions the ERC describes.

Confidential Data Plan is not a way around a prohibition. It does not make it acceptable to upload a proposal or manuscript if the funder or journal forbids that, and whether any given tool is acceptable is the funder’s or journal’s decision. When in doubt, ask them.

Conclusion

Reviewer rules come down to two ideas: the judgment is yours, and the content stays confidential. Check the policy each time, keep confidential material out of every tool, and choose tools with clear, written data terms for the language help you are allowed to use.