NIH Notice NOT-OD-25-132 states that grants containing unverifiable AI-generated content may be terminated and referred to the Office of Research Integrity. The notice does not define what “unverifiable” means. That ambiguity is the risk.
The policy is not a warning. It is active. PIs submitting R01s, R21s, and career development awards in the current cycle are subject to it. And the enforcement mechanism is not a formatting check. It is a judgment call by program officers and reviewers about whether AI-assisted content in your application can be traced back to verifiable sources.
If you used AI tools to draft any section of your grant, the question is not whether you used them. The question is whether you can prove that the output is accurate.
What the Policy Actually Says
NOT-OD-25-132 addresses the use of AI-generated text, data, and analysis in grant applications, progress reports, and peer review materials submitted to NIH. The core requirement: AI-assisted content must be verifiable. Content that cannot be verified against primary sources is treated as a scientific integrity issue, not an administrative one.
The escalation path is direct. Non-compliance does not trigger a “please revise” request. It triggers potential grant termination and referral to the Office of Research Integrity. An ORI referral is a formal investigation that appears on your institutional record and follows you to future applications, institutional affiliations, and collaborations.
The notice applies to all NIH-funded activities. That includes:
- New applications (R01, R21, R03, K-series, F-series)
- Competing renewals
- Progress reports on active awards
- Materials submitted during peer review
NIH is not policing AI use. They are policing unverifiable AI use. The difference is a verification trail.
Where the Risk Actually Lives
The highest-risk sections are the ones with the densest citation sets: the literature review, the Specific Aims rationale, and the Significance section. These are where AI-assisted drafting creates the most value and the most exposure.
Here is why. AI tools are strong at topical retrieval. They find papers relevant to your research question. They are significantly less reliable at directional accuracy, determining whether a cited paper actually supports the specific claim it is attached to. Production audits across research workflows show that 40% of AI-generated references contain errors, and only 26.5% are entirely correct.
That means a literature review drafted with AI assistance and submitted without systematic verification has a roughly 3-in-4 chance of containing at least one citation that does not say what the application claims it says.
The failure mode is specific. It is not fabricated citations (papers that do not exist). Modern AI tools produce real DOIs more reliably than they did two years ago. The failure is backwards citations: real papers, real journals, real DOIs, but findings that run opposite to the claim they are attached to. A paper cited as evidence that structured mentorship improves doctoral retention, when the paper actually found no statistically significant effect. A paper cited as supporting a drug mechanism, when the paper reported inconclusive results.
These errors survive DOI checks. They survive metadata verification. They do not survive a reviewer who reads the cited paper, or an ORI investigator who does the same.
What “Unverifiable” Means in Practice
The notice uses the word “unverifiable” without defining it operationally. Based on enforcement precedent and ORI investigation patterns, here is what it means in practice:
Unverifiable: The PI cannot produce evidence that AI-assisted content was checked against primary sources. The verification process was “I read the AI output and it seemed right.” There is no audit trail, no documented check, no evidence that any citation was independently confirmed.
Verifiable: The PI can produce documentation showing that each AI-generated or AI-assisted claim was checked against its source material. Citations were verified for existence (the paper is real), polarity (the scientific community treats the paper as supporting the stated conclusion), and alignment (the paper’s actual finding matches the specific claim).
The difference is not whether the application is perfect. Grant applications have always contained judgment calls and interpretive claims. The difference is whether there is a documented process that demonstrates due diligence in verifying AI-assisted content.
If your verification process is “I read it and it looked right,” that is not a verification record.
The Resubmission Problem
Roughly 75% of funded R01s are funded on resubmission, not on the first submission. The resubmission pathway is the normal path to funding for early-career PIs.
A termination or ORI referral on the first submission does not just lose that award. It compromises the resubmission path. Study sections remember. Program officers remember. An ORI flag on your institutional record changes how every future application is read.
The asymmetry matters. The cost of verification before submission is measured in hours and dollars. The cost of an ORI investigation is measured in years and career trajectory.
The Five-Step Pre-Submission Verification Checklist
This is what a PI should do before submitting any grant that involved AI tools at any stage of preparation.
1. Inventory AI-Assisted Sections
Identify every section where AI tools contributed to drafting, literature search, data analysis, or citation generation. This is not optional disclosure. It is the scope of your verification obligation.
2. Verify Citation Existence
For every citation in AI-assisted sections, confirm that the DOI resolves to a real, published paper with matching metadata (authors, title, journal, year). This catches fabricated references. It is the fastest check and the minimum threshold.
This step catches roughly 40% of citation failure modes. It misses the other 60%.
3. Verify Citation Polarity
For each verified citation, check how the scientific community treats the paper. Is it broadly supporting? Is it contested? Has it been retracted or corrected? Services like Scite.ai classify over 1.6 billion citation statements as supporting, contradicting, or mentioning. A paper with a high contradiction ratio cited as settled evidence is a flag that any careful reviewer will notice.
4. Verify Claim Alignment
For each citation attached to a specific claim in your application, confirm that the paper’s actual finding matches the claim. This is the check that catches backwards citations. The paper exists, the journal is real, but the finding runs opposite to what you wrote. This is the check that Reviewer #2 does manually. Run it before they do.
5. Document the Trail
Produce a verification record for each AI-assisted section. The record should show: which tool was used, what it generated, what was verified, how it was verified, and what was changed as a result. This is the document you produce if a program officer or ORI investigator asks how you ensured the accuracy of AI-assisted content.
An automated audit log is stronger than a manual attestation. A structured verification report with per-citation results is stronger than a paragraph saying “all citations were checked.”
What This Looks Like in Practice
A PI submitting an R01 with an AI-assisted literature review runs the 40 citations through a three-layer verification: existence, polarity, alignment. The system returns a structured report:
Citation [12] - Park et al. (2024)
DOI: 10.1016/j.neuroimage.2024.120482 - RESOLVED
Polarity: 67 supporting / 12 contradicting / 41 mentioning
Alignment: ALIGNED
Status: PASS
Citation [23] - Reyes & Okonkwo (2023)
DOI: 10.1038/s41593-023-01492-2 - RESOLVED
Polarity: 31 supporting / 28 contradicting / 19 mentioning
Contradiction ratio: 35.9% - CONTESTED
Alignment: PARTIAL - paper reports conditional effect, claim states absolute
Status: REFRAME - add qualification or replaceThe PI reviews the flagged citations, reframes two claims, replaces one citation, and submits. The verification report is archived. If a program officer asks, the PI produces the report showing systematic verification with per-citation results.
That is the difference between “I checked it” and a compliance record.
The Policy Is Not Going Away
NOT-OD-25-132 is one of the first formal AI verification policies from a major funder. It will not be the last. NSF, DOD, and NIH institute-specific policies are likely to follow the same pattern: not banning AI, but requiring evidence that AI-assisted work was verified.
The PIs who build verification into their workflow now are not just complying with one notice. They are building infrastructure that scales to every future policy, every funder, and every review panel that asks the same question: can you prove this is accurate?
Over 110,000 publications from 2025 are estimated to contain invalid AI-generated references. The enforcement environment is tightening because the problem is measurable and growing. Funders are responding to evidence, not speculation.
The path to compliance is not “stop using AI.” It is verification infrastructure that produces the evidence trail the policy requires. If your next R01 involves AI tools at any stage, build the verification into the timeline, not as an afterthought the night before submission.
If your grant workflow does not currently produce a verification trail, the grants pilot intake scopes what systematic verification looks like for your submission timeline and citation density. The policy is active. The next deadline is closer than the next policy update.