AI Grant Writing for Academic Researchers: What Works, What NIH Allows, and the Disclosure Trap
AI grant writing in 2026 runs into three funders with three incompatible rules. NIH treats an application "substantially developed by AI" as not the applicant's original idea. Wellcome asks applicants to disclose "substantive" AI use on the application form while exempting minimal use. Horizon Europe asks for a tool-by-tool log regardless of how much the AI shaped the text. A researcher applying to more than one of these in the same cycle cannot pick one disclosure habit and reuse it everywhere.
NIH: A Ban Framed as an Originality Standard
NIH notice NOT-OD-25-132, effective for the September 25, 2025 receipt date onward, states that NIH will not consider applications — or sections of applications — substantially developed by AI to be original ideas of the applicant (NIH Guide Notice NOT-OD-25-132). Post-award detection can trigger a referral to the Office of Research Integrity, cost disallowance, suspension, or termination. The same notice caps individual applicants at six new, renewal, resubmission, or revision applications per calendar year, a limit NIH tied directly to concern about AI-assisted mass submission. Nothing in the notice bans using AI on a grant; it bans letting AI substantially develop the ideas the application claims as original.
Wellcome: A Substantive-Use Threshold, Self-Reported
Wellcome permits generative AI in preparing an application but requires applicants to disclose any substantive use on the application form — including assisting idea generation or hypothesis development, and interpreting, analyzing, or summarizing data or text (Wellcome: Use of generative AI). Minimal use, such as translating an application into English, does not require disclosure. Wellcome states it does not use generative AI itself to assess application quality or aid funding decisions. The line Wellcome draws is about how much the tool shaped the argument, not whether it touched the document at all.
Horizon Europe: A Tool-by-Tool Paper Trail
The European Commission's Horizon Europe Standard Application Form requires applicants to list which AI tools were used, the sources used to generate or rewrite content, and an acknowledgment of the tool's limitations, including potential bias (Horizon Europe Standard Application Form). Applicants remain fully responsible for AI-generated content, and failure to disclose can render a proposal ineligible. Unlike Wellcome's substantive-use threshold, Horizon Europe's form does not carve out a minimal-use exemption — the log applies to any tool used, at any depth.
Where the Three Diverge
NIH regulates the origin of the ideas. Wellcome regulates the degree of AI involvement in producing them. Horizon Europe regulates the tools themselves, independent of degree. A researcher who documents AI use only when a funder explicitly asks will be improvising the record after the fact, because the three standards do not trigger on the same threshold. The safer default is to log every tool and every substantive step at Horizon Europe's granularity while drafting, then apply whichever funder's disclosure rule governs the specific submission.
What AI Still Does Reliably
Compliance and formatting sections, budget justification boilerplate, and cross-referencing an application against a funder's stated requirements are mechanical tasks AI handles consistently, and none of the three policies above restrict that kind of support. The risk concentrates where AI output starts functioning as the argument itself rather than the scaffolding around it.
Where Human Judgment Still Owns the Draft
The specific aims or significance section is exactly what NIH's originality standard is aimed at — the argument for why the work matters, built from a chain of reasoning a reviewer expects to trace back to the applicant, not to a model's synthesis of adjacent literature. The ICMJE's 2025 authorship rules draw a related line for manuscripts: AI can assist the writing, but a human has to own the claim. The same principle sequences a grant draft, in roughly the same four-phase order a manuscript sprint uses — structure and compliance first, argument last, and always under a human hand.
None of these three policies prohibit using AI to draft a grant application — they regulate how much of the intellectual argument it produces and how visibly that involvement gets disclosed. The reference-integrity checks in AI for Academic's Check phase catch the same citation problems in a preliminary-data section that they catch in a manuscript, before a fabricated reference becomes the reason a disclosure gets scrutinized. Free to start.