AI Grant Writing (2026): What Funders Actually Allow, a Workflow That Keeps You Honest, and a Copy-Paste Prompt

The short answer

Use AI to organize, critique and tighten your grant, never to supply its facts or its program design. As of September 2026, NIH won't treat applications "substantially developed by AI" as the applicant's original work, and NSF encourages you to disclose how AI was used and holds you responsible for accuracy. Most private foundations publish nothing, so check each RFP and ask. The workflow and the prompt are below.

Grant writing is the job nonprofits most want to hand to AI and the one where it can do the most damage. A funder is trusting that the numbers are real and the program is yours. Everything below keeps that true.

What funders actually say about AI

I'm only quoting funders from their own pages here, because summaries of funder AI policy go stale and get embellished.

NIH. Notice NOT-OD-25-132, "Supporting Fairness and Originality in NIH Research Applications," released July 17, 2025, applies to applications submitted for the September 25, 2025 receipt date and after. NIH says it will not consider applications "substantially developed by AI, or contain sections substantially developed by AI" to be the applicant's original ideas. If that's detected after an award, NIH may refer it to the Office of Research Integrity and can disallow costs, withhold, suspend or terminate the grant. The same notice limits each principal investigator to six new, renewal, resubmission or revision applications per calendar year.

NSF. Its December 14, 2023 notice on generative AI in merit review encourages proposers "to indicate in the project description the extent to which, if any, generative AI technology was used." Proposers are responsible for the accuracy and authenticity of what they submit, and the notice names fabrication, falsification and plagiarism as risks. Reviewers are prohibited from uploading proposal content into non-approved AI tools.

Private and community foundations. Most publish no AI rule. Silence isn't permission. Read the RFP and the funder's guidelines page, and if neither says anything, ask the program officer. It's a two-line email, and program officers remember the applicants who asked.

Most small nonprofits apply to foundations, not NIH. But the federal rules show where the line is drifting: AI that helps you say what you mean is fine, AI that decides what you say is not.

A workflow that keeps you honest

1. Build a fact file first. One document with your mission, program description, the numbers you can prove (people served, outcomes, cost per participant), staff bios, budget summary and your best past proposals. This is the only source AI is allowed to use. Put it in a shared project so every grant starts from it; how to use Claude Projects shows the setup.
2. Read the funder's rules. The RFP, the scoring criteria, and any stated AI policy. Paste the criteria in word for word.
3. Map criteria to facts. Ask AI to list each criterion and which facts in your file answer it, and to flag every criterion your file can't support. The gaps are the most useful output of the whole process.
4. Write the heart yourself. The need, the program design, the story of one participant. Rough is fine. Then let AI restructure and tighten it to the word limit.
5. Run a reviewer pass. Have AI score your draft against each criterion, as a skeptical reviewer would. OpenAI's own nonprofit help article suggests exactly this: upload the draft and the donor's criteria and ask it to assess the fit. The prompt is below.
6. Check every number. Every figure, date, citation and name in the final draft gets checked against the fact file by a person. Anything not in the file comes out.
7. Disclose if asked. If the funder wants to know, say how you used AI: organizing, editing and critique, with all facts and program content supplied by staff.

The copy-paste reviewer prompt

Paste this with your draft, the funder's criteria and your fact file attached:

You are a skeptical grant reviewer for [FUNDER NAME].

Attached: (1) the funder's scoring criteria, (2) our draft proposal, (3) our fact file.

For each criterion:
1. Score the draft 1-5 and explain the score in two sentences.
2. Quote the sentence in the draft that best answers it, or say "not addressed".
3. Name the one change that would raise the score the most.

Then:
- List every number, date, name or citation in the draft that does NOT appear in the fact file. Do not correct them. Just list them.
- List any claim a reviewer would ask us to prove.
- Do not add new facts, outcomes, statistics or sources. If something is missing, say what we need to supply.

The last three lines matter most. They turn the AI from a writer into an auditor, which is the job it's safest at in a grant.

Which tool, and where the data goes

Use a business plan, not a personal account. Your fact file has donor, participant and budget detail in it. ChatGPT Business and Claude Team both say they don't train on your content by default, and both are $8 a user a month for eligible nonprofits. The details are in ChatGPT for nonprofits and Claude for nonprofits; Claude's nonprofit plan also shows a Candid connector, which is where funder research starts.

Write the rule down so it survives staff turnover: approved tool, what never goes in, and that a person checks every fact before a proposal leaves the building. My one-page AI acceptable use policy covers it with one extra line for grants. If you're deciding where else AI fits in a small nonprofit, the wider picture is in AI for nonprofits.

Last verified: September 27, 2026, against grants.nih.gov (NOT-OD-25-132), nsf.gov and help.openai.com. Funder policies change; check each funder's own page before you submit.