AI Grant Disclosure Rules in 2026: NIH, Wellcome, and Horizon Europe Side by Side
NIH AI disclosure sits in an awkward place: the agency's policy is a restriction, not a checkbox. Wellcome asks for a declaration on the application form. The EU leans on grant-agreement obligations rather than an application question. A researcher holding applications to more than one of these funders needs to know which rule governs which submission, because they do not agree on where disclosure goes or what triggers it.
NIH: a restriction, with the PI accountable
On 17 July 2025, NIH issued notice NOT-OD-25-132, which states that it "will not consider applications, or sections of applications, substantially developed by AI to be the original ideas of the applicant" (NIH Grants). Assistive use is not banned. What is missing is any disclosure field that makes substantial AI drafting acceptable — the applicant owns the content regardless of what produced the first draft.
The same notice caps a principal investigator at six new, renewal, resubmission, or revision applications across all council rounds in a calendar year, applied without exception from the start of 2026 (NIH Extramural Nexus). The two rules share a rationale: applications are supposed to be the applicant's own ideas, submitted in deliberate numbers.
Wellcome: declare substantive use on the form
Wellcome permits generative AI in preparing an application but requires applicants to tell it about "any substantive use of generative AI on their application form" (Wellcome: Use of generative AI). Its examples of substantive use include idea generation, developing hypotheses, and interpreting, analysing, or summarising data. Minimal use, such as grammar correction, does not need to be declared.
This is a positive obligation. Silence where AI was used substantively is a policy breach, not a neutral omission. The disclosure lives on the form itself, so it has to be decided before submission, not added later.
Horizon Europe: the obligation is on the beneficiary
The EU does not run an application-stage AI question comparable to Wellcome's. The reference point is the European Research Area Forum's 2024 guidelines on the responsible use of generative AI in research, which place accountability on the researcher: use the tools transparently, indicate which were used and how, list sources including AI-generated ones, and monitor for plagiarism (European Commission). Evaluators, separately, are advised not to put proposals into generative AI tools.
Those expectations do not vanish at award. They carry into the grant agreement's confidentiality, integrity, and intellectual-property obligations for the funded project, which is where an audit would look.
The operating rule for a mixed portfolio
The three regimes disagree on mechanism but agree on one point: the human applicant owns every sentence, and disclosure does not transfer that. A rule that satisfies all three:
Keep a running log as you draft — which tool, which section, what it did — recorded at the time, not reconstructed afterward. Then disclose to the strictest standard any funder in your portfolio requires. If one application in a cycle goes to Wellcome, write the declaration for all of them. It costs nothing on the funders that do not ask, and the record exists if a question comes up later. This is a recommendation, not a policy requirement, but it is cheaper than reconstructing provenance under pressure.
What none of them accept
No funder here treats "AI wrote it, and that was disclosed" as sufficient for the intellectual core: the aims, the hypothesis, the significance argument. Disclosure covers process. NIH's notice is explicit that AI-developed ideas are not the applicant's own, and the same principle runs through the ICMJE authorship rules for manuscripts. The parallel with keeping a visible record of the writing process, rather than a binary claim, is the same shift described in proving your AI writing process. For the wider view of what AI does well and badly in grant text, see AI grant writing for academic researchers.
AI for Academic's workspace at aiforacademic.world keeps each tool output as a saved artifact per project, which is a starting point for the AI-use log these policies now expect. Free to start.