Using AI to Write NIH Grants in 2026: What Works, What Crosses the Line

What the 2026 Nature Study Actually Found (and Didn't)

The study that set off the current round of arguments looked at roughly 1,400 R01, R21, and K-series applications submitted between late 2023 and mid-2025, cross-referenced with survey data on whether the PI used a large language model during drafting. The headline finding was a modest but statistically meaningful bump in funding probability for AI-assisted applications — a few percentage points at most, not the dramatic swing that got turned into a headline.

Three things the study did not show get lost in the discussion. It did not show that AI wrote the proposals; in every case the PI wrote the scientific content, and AI was used for editing, restructuring paragraphs, or generating alternative phrasings. It did not separate out which sections benefited; most of the lift was in the Approach and Significance sections, not the Specific Aims page. And it did not track which reviewers flagged AI-like language in critiques, which is the number I would most want to see.

There is also a selection problem that the authors acknowledge in the supplementary text. PIs who adopt new tools early tend to be, on average, more organized and more responsive to feedback than PIs who do not. Some of the "AI effect" is probably just a "PIs who iterate on drafts" effect. You can get the same lift by showing your draft to three outside readers, which is what most experienced writers already do.

Where NIH Policy Stands in April 2026

NIH banned reviewers from using generative AI tools on applications back in NOT-OD-23-149, and that ban is still in place. What changed in the last year is the applicant side. The February 2026 guide notice clarifies that AI use by applicants is not prohibited outright, but content "substantially generated" by a language model without meaningful human revision is considered a research integrity concern if submitted as the PI's original writing. The notice also tightens expectations on the Biosketch and Specific Aims page — the sections most likely to show up verbatim in future applications.

In plain language: you are allowed to use AI as a writing assistant. You are not allowed to submit text that a model produced on its own. If asked, you should be able to explain what you did and show your drafts. That second part matters more than it sounds. A few program officers I have spoken with have begun asking, informally, whether an applicant's Aims page sounds like them. The question is asked casually but it is not a casual question.

The practical rule I give first-time R01 writers

Do not paste a single sentence from a model into a section of your application without rewriting it. Use the model to ask questions, surface alternatives, and catch issues. Write the actual text yourself. The sentences in the final PDF should come out of your own keyboard.

The Homogenization Problem Reviewers Are Starting to Flag

Talk to experienced study section reviewers and a pattern comes up. Applications that lean too hard on AI assistance start to sound alike. They use the same transition phrases, the same rhythm of topic sentences, the same default hedges. "These findings will inform future therapeutic strategies" is a sentence that appears, or some close variant of it, in somewhere between a third and half of recent applications. Reviewers notice. The word I have heard used more than once is "beige."

Reviewers also notice the opposite problem: an Approach section written in a flat, perfectly grammatical, slightly stiff register, followed by a paragraph in the Aims page that sounds like a different person entirely. Voice inconsistency reads as disengagement with the writing, which in turn reads as disengagement with the science. It is not a fatal flaw, but it is a foot on the scale in the wrong direction.

The fix is not to avoid AI. The fix is to stop treating AI output as the writing and start treating it as the draft. If the voice is uniformly yours — which takes a pass of manual rewriting, not just a "make it sound more natural" prompt — the homogenization problem disappears. Reviewers are not trying to catch you using a tool. They are catching the loss of voice that happens when you use the tool badly.

Six Places AI Genuinely Earns Its Keep

Here are the uses I have seen actually improve proposals, with no ethical ambiguity and no policy risk.

1. Stress-testing your Specific Aims logic

Paste your draft Aims page and ask: "What questions would a reviewer in an adjacent subfield raise about this page?" The answers will not be perfect, but they will flag three or four things you stopped seeing after your tenth re-read. Treat it as a pre-review, not as edits to accept.

2. Finding every passive construction in a draft

Ask the model to list every passive-voice sentence in a paragraph. Rewrite them yourself. This is the cleanest "AI as grep" use case — the model is faster than you at finding the issue, and you are faster at fixing it than it is.

3. Outlining Approach sections before you write them

Feed the model your Aims and ask for a section-by-section outline for the Approach. You will almost certainly disagree with half of it. That disagreement is the useful output — it surfaces the structural choices you are making without noticing, and lets you make them deliberately.

4. Checking that rigor language is covered

Ask: "Does this Approach section address scientific premise, authentication of key resources, and sex as a biological variable?" If the answer is "partially," you know exactly what sentence to add.

5. Reformatting references and checking style

AI is excellent at catching NIH-style formatting inconsistencies, spotting missing first names in a bibliography, and flagging where your section headings drift from the Funding Opportunity Announcement's language. This is tedious human work that AI is genuinely better at.

6. Generating "worst-case reviewer" critiques

Ask the model to write the harshest plausible critique of your Significance paragraph, as if it were a reviewer who does not buy your premise. This gets you the objections your friendly colleagues are too polite to raise. Address them in the draft, not in the rebuttal letter.

Four Places It Will Quietly Sink Your Score

Writing the Specific Aims page

The Aims page is the one page every reviewer reads carefully. It is also the page where model phrasing is most recognizable. I have seen applications where the Aims page reads as distinctly different from the Research Strategy, and reviewers comment on it. If nothing else in your application comes from a model, let it be this page.

Summarizing your own preliminary data

The model does not know what your data mean. It can produce plausible-sounding summaries that miss the one nuance you wanted reviewers to catch. This is one of the few places where AI-generated prose actively misrepresents the science, not just smooths it.

The Biosketch Personal Statement

The Personal Statement is a place for voice. Model-drafted personal statements tend to sound competent and generic, which is exactly the impression you do not want to leave in the Investigator score. Write these paragraphs yourself, even if the rest of the biosketch was assembled with help.

Literature review paragraphs

Do not use AI to generate citations or summarize the literature. Models still hallucinate references at a rate that is embarrassingly high, and reviewers who catch a phantom citation will distrust the rest of your proposal on principle. A wrong citation in a grant is a self-inflicted wound.

A Workflow That Keeps the Application Defensible

Here is the approach I recommend to first-time R01 writers who want to use AI responsibly without crossing lines.

Write the first draft of the Aims page by hand. No AI involvement. This is the document the rest of the application is fitted to. Its voice has to be yours from the start.
Use AI as a reviewer on your draft Aims. Ask for objections, logical gaps, and ambiguities. Rewrite in response. Do not paste the model's rewrites.
Outline the Research Strategy yourself, then ask AI to critique the outline. Again, the model's best work is pointing out what you have left out, not what you should write.
Draft each Approach subsection by hand. Only after a full human draft exists should AI come in for editing passes.
Use AI for targeted edits: passive voice, sentence length variation, FOA alignment, rigor language. These are the edits where the model is cleanly better than you at finding the issue, and you are cleanly better at fixing it.
Read the final application aloud. Fix every sentence that sounds like it did not come from you. This is the only check that actually catches homogenization. Reading on screen is not the same as reading aloud.

If you ever have to answer a question about how AI was involved in the proposal, this workflow gives you an honest, specific answer. "I used AI to stress-test my arguments, check for passive voice, and align my section headings to the FOA. I wrote every sentence in the application myself." That is a defensible answer. The alternative — "I had the model draft the Approach and then I edited it" — is not.

On Disclosure and the CPOS Common Form

As of April 2026, NIH does not require disclosure of AI assistance for drafting. That may change. The Common Project Object Service (CPOS) Common Form rollout, which has been phased in across 2025 and 2026, has added language around research integrity that many institutions are interpreting to cover AI usage in proposal preparation. A number of universities — I have heard of at least seven — have begun asking PIs to document AI tool use during internal proposal review, separate from NIH's own requirements.

The safe assumption is that disclosure is coming. Keep a short log of which tools you used for what, on which sections, during drafting. If the policy lands mid-cycle, you will be glad you have that log. More importantly, the act of keeping the log is itself a check on whether your usage has drifted from assistance to generation.

Frequently Asked Questions

Can I use AI to translate my proposal from a non-English first draft?

Yes, this is widely accepted. Non-native English speakers have used translation and editing tools long before generative AI, and NIH's policy does not treat translation as generation. That said, follow the translation with a careful read-aloud pass — translated prose can land in a tone that does not sound like you even when every word is accurate.

Is it safe to run my Aims page through an "AI humanizer" tool?

No. Those tools tend to produce distinct artifacts of their own — oddly varied sentence lengths, unusual word choices — that reviewers are starting to recognize. If the page needs to sound more human, rewrite it by hand.

My institution provides a "secure" AI tool. Does that change the policy answer?

Not meaningfully. The question NIH cares about is whether the PI wrote the proposal. Whether you used a commercial model or a private deployment, the rule is the same: AI can help, but the sentences must be yours.

What about AI for the Data Management and Sharing Plan?

The DMS Plan is lower risk than scientific sections because the content is largely procedural. Using AI to generate a first draft of the plan, based on your inputs about data types and repositories, is common practice and not a policy issue. You still need to read it carefully and make sure the plan matches what you will actually do.

Does the Nature study mean I should start using AI on my next application?

Only if you were not already iterating on drafts. The study's finding is consistent with "more editing rounds produce better applications." If you already do three rounds of revision with outside readers, AI will add less. If you do not, AI is a cheap way to get one additional round of feedback without bothering a colleague.