Grant Writing and AI: Disclosure Expectations From Funders

A program officer reads a proposal that’s unusually smooth, then notices a cited study that doesn’t exist. Now the whole application is suspect, and a reviewer who used a chatbot to “tighten the prose” has a much bigger problem than clunky sentences. Grant writing is one of the places where AI can quietly help and quietly sink you, and the difference often comes down to one thing: what you disclosed, and to whom.

Key takeaways

• Funder rules vary a lot. Some bar AI from peer review entirely, some ask for an AI-use statement, and many are still silent. Read the specific solicitation before you assume anything.
• Confidentiality is where people get into real trouble. Pasting an unfunded proposal, preliminary data, or someone else’s application into a public AI tool can breach reviewer agreements and institutional data rules.
• Disclose AI the way you’d disclose any drafting help a funder would reasonably want to know about, and name what the tool actually did instead of hiding behind vague words like “AI-enhanced.”
• You stay accountable for every claim, number, and citation. A model can invent a reference or a statistic with total confidence, so verification isn’t optional.
• Don’t lean on AI detectors to settle this. They output probabilities, misfire on non-native English writers, and can’t prove who wrote what.

Why grant writing is its own disclosure problem

Most “should I disclose AI?” debates are about credit and honesty. Grant writing adds two things that raise the stakes: money and confidentiality.

The money part is obvious. A funder is deciding whether to hand over real resources based partly on the words in front of them. If those words misstate what’s been done, overstate preliminary results, or cite work that isn’t real, that’s not a stylistic slip. It can read as research misconduct, and it can follow you and your institution.

The confidentiality part is the one people underestimate. Proposals contain unpublished ideas, draft aims, and sometimes data nobody outside the team has seen. When you serve as a reviewer, you’re handed other people’s confidential applications under an explicit agreement. Drop any of that into a public AI tool and you may have shipped protected material to a third party that could use it to train models. That’s why the disclosure question here isn’t just “did you mention it?” It’s also “was it even allowed to leave your screen?”

What funders are actually saying

There’s no single rule, so the only safe move is to check the specific funder and the specific solicitation. That said, a pattern has emerged.

On the reviewer side, several major funders have drawn a hard line. In a June 2023 notice (NOT-OD-23-149), the U.S. National Institutes of Health barred scientific peer reviewers from using generative AI tools to analyze and formulate critiques of grant applications, citing confidentiality. The National Science Foundation followed in December 2023, telling reviewers not to upload proposals or any portion of them to non-approved generative AI systems. The logic is consistent: a confidential application can’t be fed to an external tool, full stop.

On the applicant side, the picture is looser and still moving. Some funders now ask applicants to disclose AI use or include an AI-use statement, and NSF has gone as far as encouraging proposers to mark AI-generated elements. Others permit AI assistance but stress that the applicant remains fully responsible for the content, errors and all. Many haven’t published anything specific, which is not the same as a green light to hide it. Silence means you fall back on the general principles of honesty and accountability, not that nobody cares.

The practical reading

If you take one habit from this article, make it this: before you open any AI tool for a grant, open the funding opportunity announcement and the funder’s policy page first. Look for the words “artificial intelligence,” “generative AI,” or “large language model.” Then look at your institution’s research-office guidance, which may be stricter than the funder’s. The rules that bind you are the strictest applicable layer, not the most convenient one.

What needs disclosing, and what doesn’t

Not every use of software needs a note. Running spell-check or a grammar pass over text you wrote is the kind of help nobody expects you to flag. The honest test is the same one that works everywhere else: would knowing change how a reviewer judges the reliability or originality of your work?

By that test, lean toward disclosing when AI:

• Drafted substantive narrative a reviewer assumes you wrote, such as the significance, approach, or innovation sections.
• Generated or summarized analysis you present as your own reasoning.
• Produced text, figures, or a literature summary that shapes how the science reads.
• Wrote anything you didn’t independently verify line by line.

You generally don’t need to flag minor copy-editing, reference formatting, or transcription of your own notes. When in doubt, disclosing costs you little and protects you a lot.

A line to avoid

Steer clear of “AI-enhanced” and similar marketing-flavored phrases. They tell a reviewer nothing. “Enhanced” could mean a comma fix or a fully synthetic Specific Aims page, and that gap is exactly the thing disclosure is supposed to close. Vague disclosure is barely better than none.

You are still the author of every claim

Here’s the part no policy can do for you. A language model produces fluent, confident text whether or not it’s true. It will write a plausible-looking citation for a study that was never published. It will state an effect size it has no basis for. It does this without malice and without any signal that it’s guessing.

In a grant, that failure mode is brutal. Reviewers are experts in your field. A fabricated reference, a misattributed finding, or a statistic that doesn’t match the cited source is the kind of thing they catch, and once they catch one, they stop trusting the rest. So treat every AI-produced fact, number, and citation the way you’d treat an anonymous tip: possibly useful, definitely unconfirmed until you check it against the primary source yourself.

Mini-scenario: the “tightened” methods section

A postdoc is up against a deadline on an NIH R01 resubmission. To save time, she pastes the full draft, including unpublished preliminary data, into a public chatbot and asks it to tighten the approach section. The prose comes back cleaner. It also smooths two hedged statements into firm claims and adds a citation that sounds right but points to a paper that doesn’t exist.

Three things went wrong, and only one is really about writing. The confidential, unpublished data left her control and went to a third-party tool, possibly against her institution’s policy. The model overstated results she’d deliberately hedged. And it inserted a fake reference a reviewer would likely flag. None of these were caught because the output “read well,” and reading well is precisely what a model is good at even when it’s wrong. Fix the process and the same tool turns harmless: use an approved tool, keep confidential data out, and verify every line.

How to write the disclosure

Keep it short, specific, and honest. A useful AI-use statement names the tool, says what it did, and confirms a human verified the output and takes responsibility. Something like:

> “We used a large language model to draft and edit portions of the project narrative. All scientific claims, data, and citations were written and verified by the named investigators, who take full responsibility for the content.”

Put it where the funder asks. If the application has no designated spot, a brief note in the cover letter or the relevant section is reasonable. The goal is for a reviewer to feel informed, not ambushed.

Don’t let an AI detector run the conversation

It’s tempting to imagine funders quietly running submissions through an AI detector, and to write defensively because of it. Don’t. Detectors don’t tell anyone the truth about how text was made; they output a probability, and they’re wrong often enough to be dangerous as evidence. A 2023 Stanford study found several detectors disproportionately flagged writing by non-native English speakers as AI-generated, which matters in a field full of international researchers. OpenAI retired its own classifier in July 2023, citing low accuracy. If you want the mechanics, we’ve explained why AI detectors output probabilities, not proof.

A detector score can’t confirm whether you used AI, and it can’t confirm whether you disclosed honestly. Your real protection is the same on both sides of the table: accurate, verified content and clear disclosure. For more on how detection actually behaves, you can dig into more of our writing on AI, writing, and detection.

Frequently asked questions

Do I have to disclose AI use in a grant proposal?

It depends on the funder, so check the specific solicitation and the funder’s policy page first. Some funders now require an AI-use statement; many are still silent, which is not the same as permission to hide it. The honest test is whether knowing would change how a reviewer judges your work. If AI drafted substantive narrative, generated text you present as your own analysis, or shaped the science, lean toward disclosing it clearly. Routine spelling and grammar help generally doesn’t need a note.

Can I paste my draft proposal into ChatGPT to improve it?

Be careful. Your own early-stage ideas may be fine, but unpublished aims, preliminary data, and anything covered by confidentiality or institutional data rules can be a real problem once it leaves your control. Check your institution’s policy and the funder’s terms, and prefer enterprise or institutionally approved tools that don’t train on your inputs. Never paste a proposal you’re reviewing for a funder into a public AI tool, because several funders explicitly forbid that.

Is it against the rules to use AI in peer review of grants?

For some major funders, yes. The NIH has barred reviewers from using generative AI tools in the review of grant applications, citing confidentiality, and the NSF has told reviewers not to upload proposals to non-approved generative AI systems. The concern is that uploading a confidential application to an external tool breaches the reviewer’s obligations. If you serve as a reviewer, treat the application as confidential and follow the funder’s review instructions exactly.

Will a funder run my proposal through an AI detector?

Some might, but it’s a weak basis for any decision and you shouldn’t write in fear of it. Detectors output probabilities, not proof, and they misfire often. A 2023 Stanford study found several detectors disproportionately flagged writing by non-native English speakers as AI-generated, and OpenAI retired its own classifier in July 2023 for low accuracy. A detector score can’t confirm whether you used AI or whether you disclosed it honestly. Your protection is accurate work and clear disclosure, not a low score.

How do I write an AI-use statement for a grant application?

Keep it short, specific, and honest. Name the tool, say what it did, and confirm you verified the output. For example: “We used a large language model to draft and edit portions of the narrative. All scientific claims, data, and citations were written and verified by the named investigators, who take full responsibility for the content.” Put it where the funder asks; if they don’t specify, a brief note in the cover letter or the relevant section is reasonable.

The bottom line

AI in grant writing isn’t an ethics problem because a tool helped you draft. It turns into one when confidential material leaks, when nobody verified the claims, or when a reviewer finds out about the AI’s role only after the fact. Get the order right: check the funder’s rules, protect confidential data, verify every number and citation, then disclose plainly. When your last step is making the writing sound like you, you can make an AI-assisted draft read like you actually wrote it and submit something you’re glad to put your name on.