The NIH just dropped a bombshell: AI-written grant proposals are officially out. Starting with the September 25, 2025 deadline, any NIH application that’s “substantially developed by AI” will be considered ineligible. If they catch it after an award’s made? You could lose the grant entirely.
Let’s unpack what this means, why it matters, and how to stay out of trouble.
What the Policy Actually Says
The new guidance (NOT-OD-25-132) comes from NIH’s Office of Science Policy. It outlines two big changes:
-
No more AI-generated applications: Proposals that are “substantially developed by AI” won’t be considered original work. If detected post-award, NIH may report it as research misconduct and yank your funding.
-
New submission cap: NIH will now only accept six new, renewal, resubmission, or revision applications per Principal Investigator per calendar year. This includes anyone listed as a multiple PI.
NIH says they’ve seen investigators use AI tools to submit more than 40 applications in a single cycle. Honestly, I’m impressed. But you can see why they’re tightening the reins.
Why NIH Took This Step
According to NIH, the rise of AI tools like ChatGPT has made it way too easy for investigators to crank out low-effort, high-volume applications. That’s flooding peer review panels, compromising quality, and stretching administrative capacity.
It’s not just about fairness; it’s also about integrity. NIH flagged the risk of plagiarism, fabricated citations, and the fact that AI can generate text that sounds confident but is flat-out wrong. If you’re running a serious research program, that’s not a gamble you want to take.
So… What Does “Substantially Developed by AI” Mean?
Here’s the tricky part: NIH didn’t define exactly where the line is. But based on years of proposal writing experience, here’s a good rule of thumb:
-
Allowed (probably): Using AI to brainstorm bullet points, rephrase awkward sentences, or summarize literature (if you verify the sources).
-
Risky: Asking AI to write your Specific Aims or draft a full Background section.
-
Not allowed: Submitting anything that reads like it came out of a proposal-writing machine without serious human oversight.
If your proposal’s voice, logic, or innovation isn’t clearly yours, you’re walking into dangerous territory.
Can NIH Actually Detect AI-Written Text?
Here’s where things get murky. NIH says it “will continue to employ the latest technology in detection of AI-generated content.” But here’s the truth: there are currently no reliable tools that can consistently detect AI writing with high accuracy. Multiple peer-reviewed studies and real-world tests have shown that AI detectors often:
-
Flag human-written content as AI (false positives).
-
Miss AI-generated content that’s been lightly edited (false negatives).
-
Struggle most with technical or scientific writing, which often overlaps stylistically with AI output
In short, enforcement will be tough. But don’t mistake that for a free pass.
This policy isn’t just about catching violators—it’s about shifting the culture. NIH is making it clear: use your brain, not just your tools. Even if detection tech is shaky now, it’s improving fast. And worse than being flagged by an algorithm is being caught because your proposal just doesn’t hold up under review.
What This Means for SBIR/STTR Applicants
For academic investigators, the six-application cap could discourage participation in SBIRs, especially as a Multi-PI on Phase I proposals where the PI’s contribution is smaller. Why burn one of your six chances on a $305K subaward?
For startups using AI-heavy proposal writers (or consultants who rely on them behind the scenes), the risk just shot way up. NIH is watching, and they claim they’ve got detection tools.
And for agencies like NSF and DOE? They’re probably taking notes. I wouldn’t be surprised if similar policies show up elsewhere soon.
Proposals Need to Be Human
Honestly, I think this is a good thing. At GrantCraft, we’ve explored proposals that clearly came straight from bots. Sure, the grammar’s fine. But strong proposals are built on clarity of thought, strategic alignment, and a deep understanding of the science, the reviewer’s mindset, and the funding agency’s mission. AI can’t fake that (at least not well).