NIH AI Grant Policy NOT-OD-25-132: What 'Substantially Developed by AI' Means

The most-cited sentence in NIH's AI policy is also its most underspecified. "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." Ten months after the notice dropped and eight months after it took effect, every research office in the country has read that sentence dozens of times — and few can tell a principal investigator with certainty where "substantially" begins. What has emerged in practice is not the framing most reporters covered when the policy first appeared.

What NOT-OD-25-132 Actually Says

NIH issued NIH Notice NOT-OD-25-132, "Supporting Fairness and Originality in NIH Research Applications," on July 17, 2025, with publication in the Federal Register on August 5 and an effective date of September 25, 2025 for all council rounds. The notice does two things that often get conflated. First, it caps each principal investigator at six new, renewal, resubmission, or revision applications per calendar year across all council rounds, a response to application-volume pressures NIH attributed in part to AI-assisted submissions. Second, it states that AI tools "may be appropriate to assist in application preparation for limited aspects or in specific circumstances," but that applications "substantially developed by AI" will not be treated as the applicant's original ideas. Detected post-award use of AI may be referred to the Office of Research Integrity, with enforcement options including disallowing costs, withholding future awards, suspension, and termination.

The policy is not a ban. It is an originality standard with consequences.

The Line NIH Did Not Draw Explicitly

"Substantially" carries the weight in the notice, and NIH chose not to define it. There is no word-count threshold, no percentage cutoff, no per-section rubric. Research offices have read this absence two ways. The cautious read is that any AI involvement above the level of grammar-checking is risky, and the safest posture is to disclose minimally and use AI minimally. The pragmatic read — increasingly the dominant one inside university sponsored-programs offices — is that NIH is signaling a test that focuses on whether the ideas are the applicant's, not on whether the tools touched the document.

This pragmatic read tracks a distinction the notice itself implies but never names: the difference between AI as a disclosed collaborator and AI as an undisclosed ghostwriter. A disclosed collaborator helps a researcher articulate their own scientific reasoning, polish prose, organize a complex aims page, or stress-test a hypothesis. An undisclosed ghostwriter generates the scientific reasoning itself and lets a human submit it under their own name. NIH's enforcement concern, and ORI's interpretive lens, is the second case.

Other federal funders have moved in similar directions. NSF and the Department of Energy's Office of Science have signaled that disclosure-based frameworks are the likely landing point for AI in application preparation, even as both agencies have separately declined to authorize AI use inside peer review. Most major private foundations have not yet published standalone AI policies, in effect deferring to whatever standard the federal funder of record applies.

What AI Use Is Actually Safe Under the Policy

The list below reflects the working consensus across sponsored-programs offices, not NIH guidance. Confirm specifics with your institution's research integrity office before submission.

• Uncontroversial: grammar checking, citation management, formatting, accessibility checks, reference list assembly.
• Broadly accepted: helping a non-native English speaker polish prose without changing scientific content.
• Acceptable with internal documentation: AI-assisted literature review in which the researcher verifies every citation and writes the synthesis themselves.
• Gray zone: AI drafting of background or significance sections that the PI then substantively rewrites. The working test is whether a reviewer reading the rewritten section would recognize the PI's own voice and reasoning.
• Clearly proscribed: AI-generated specific aims pages; AI-generated research strategy without human-original thinking; fabricated citations of any kind; AI-generated text presented as if it represented the PI's reasoning when it did not.

As of May 2026, ORI has not issued a public referral specifically tied to NOT-OD-25-132 — but ORI has the authority to act.

How Research Offices Are Advising Principal Investigators

The de facto disclosure norm that has emerged across many large research universities is short and operational. Internally document which sections involved AI assistance and at what level. Keep a record of the prompts used and the model versions. Treat AI like any other writing assistant: human author, human responsibility, human signature. Do not, under any circumstance, submit AI-generated text that the PI cannot defend in conversation with a program officer.

This is not the posture most coverage of the policy predicted. The early framing in mid-2025 leaned toward "NIH bans AI in grant applications," and a vocal segment of the research community read the notice as the end of legitimate AI use in proposal preparation. The actual settling point, ten months in, is more practical and more enforceable: NIH does not care what tools you used. NIH cares whether the ideas are yours. The originality test, not the tool test, is the rule.

Why the Originality Test Is Spreading

The disclosed-collaborator framing is spreading across federal agencies. A categorical ban is unenforceable at scale, and a tool-detection arms race is unwinnable. An originality standard, by contrast, is a standard human reviewers have been applying for fifty years. AI use becomes a question of authorship and disclosure rather than a question of contraband.

For researchers writing applications today, the operational implication is direct: use AI in a way you can defend if a program officer asks. That is the test Granted's AI grant writing platform is built around — keeping the researcher's reasoning at the center while the tool handles the surrounding work.