Grant writing was one of the first places researchers reached for a language model, drafting boilerplate, tightening a specific aims page, restructuring a budget justification. Funders noticed, and over the course of roughly a year, the three largest funding bodies a working researcher is likely to deal with landed on three genuinely different policies. None of them agree on what "using AI" in a proposal should mean, or how much of it is acceptable, and getting it wrong under the strictest of the three doesn't cost you a stern email. It costs you a proposal that never gets reviewed at all.
NIH: a ban dressed as an originality requirement
NIH's policy, Notice NOT-OD-25-132 ("Supporting Fairness and Originality in NIH Research Applications"), took effect for applications due September 25, 2025 onward. It does not offer a disclosure box to tick.
"NIH will not consider applications that are either substantially developed by AI, or contain sections substantially developed by AI, to be original ideas of applicants."
NIH Notice NOT-OD-25-132
There is no defined threshold for "substantially," which means the policy functions less like a rule with a bright line and more like a standing risk that any given proposal could later be judged to have crossed one. The notice also spells out what happens if that judgment comes after the money has already gone out:
"If the detection of AI is identified post award, NIH may refer the matter to the Office of Research Integrity to determine whether there is research misconduct while simultaneously taking enforcement actions,"
up to and including "disallowing costs, withholding future awards, wholly or in part suspending the grant, and possible termination" (NIH Notice NOT-OD-25-132).
NSF: disclosure encouraged, misconduct redefined
NSF took the opposite posture on disclosure itself:
"Proposers are encouraged to indicate in the project description the extent to which, if any, generative AI technology was used and how it was used to develop their proposal."
NSF, official notice to the research community on AI
Encouraged, not required. Where NSF did tighten its language was research misconduct: fabrication, falsification, or plagiarism "committed by an individual directly or through the use or assistance of other persons, entities, or tools, including artificial intelligence-based tools" is explicitly prohibited in proposing, performing, reviewing, or reporting on NSF awards. Reviewers get a harder rule than applicants do: "NSF reviewers are prohibited from uploading any content from proposals, review information and related records to non-approved generative AI tools" (NSF notice to the research community on AI).
Horizon Europe: declare it, and be ready to source it
The European Commission's Standard Application Form now includes explicit generative AI language. Applicants:
"...are fully responsible for parts that are produced by an AI tool," "must be transparent concerning the tools they used," and "must be able to provide a list of sources used to generate/rewrite content and citations."
Horizon Europe Standard Application Form, European Commission
That last clause is easy to miss and hard to satisfy after the fact: providing "a list of sources used to generate/rewrite content" assumes someone kept track of which sources fed which paragraph while the proposal was being written, not after (Horizon Europe Standard Application Form; RVO.nl guidance for applicants).
The same three funders, side by side
| Funder | Stance on AI use | Where it is declared | What triggers a penalty |
|---|---|---|---|
| NIH | Applications "substantially developed" by AI are not reviewed | No disclosure mechanism; the bar is originality, not declaration | AI-caused fabrication or plagiarism → possible Office of Research Integrity referral |
| NSF | Use is permitted; disclosure is encouraged, not mandatory | Project description (optional) | Fabrication, falsification, or plagiarism via any tool, AI included |
| Horizon Europe | Use is permitted and expected to be declared | Standard Application Form | Non-disclosure, not the AI use itself |
Three policies, one shared blind spot
Whatever their philosophy, all three funders are asking a PI to self-report something they have no independent way to verify. NIH's "substantially developed" threshold is judged after the fact, by reviewers who were not present while the proposal was written. NSF's "encouraged" disclosure and Horizon Europe's mandatory declaration both rest on the applicant's own account of a process nobody else observed. This is the same structural gap we have written about for authorship disclosure generally: a self-report is not evidence, and it fails honest and dishonest applicants in exactly the same way, because there is no mechanism to distinguish them.
Turning a self-report into a number
Write a proposal in TeXposit and that self-report stops being the only account of what happened. The provenance report attached to a project gives Horizon Europe's declaration a specific, word-level figure to point to instead of a checkbox: which sections a human typed directly, which came from an accepted AI suggestion, and how much of the applicant's own prose statistically correlates with the proposal's AI chat log without having been inserted by it. For NIH's ban, the same report is what a PI would want on hand if a program officer ever questioned whether a proposal was, in fact, substantially AI-developed.
It doesn't certify compliance with any of the three. That judgment belongs to the funder, applied to the finished proposal. What changes is what a PI has to offer when asked: not a recollection of how the thing came together months later, but a record that was already sitting there.