AI for grant writing: what 8 funders allow, and 8 prompts from RFP to draft

Can you use AI for grant writing?

Yes, and NIH is the only one of the eight funders read for this page on 3 October 2026 that says it runs AI detection on applications. All eight permit some AI help in preparing an application. They split on two questions: whether you have to say so, and what happens when the AI supplied the ideas. Two expect a declaration (UKRI and Lloyd's Register Foundation), NSF and the Wenner-Gren Foundation encourage one, the ARC wants it only when it asks, and three don't ask.

The table gives each rule as written, with the reviewer side, which matters for the paste question further down. Dates are the policy's own issue or revision date. Every row was read at the funder's page on 3 October 2026 (NIH's through the Internet Archive copy of 24 September 2026, because grants.nih.gov blocks automated reads).

Seven of the eight forbid their own reviewers from feeding your proposal into a public chatbot. That's a fair measure of how funders rate the confidentiality of the tools applicants draft in. Private foundations mostly have no written rule. The Center for Effective Philanthropy's survey was published 30 September 2025. In it, 61% of nonprofits that use AI used it for development and fundraising, grant applications included (279 organizations). Yet fewer than 20% of nonprofit leaders said any funder had discussed AI with them. Where a foundation is silent, look for an AI question on its application form and ask the program officer before you decide what to disclose.

Can you use AI for NIH grants?

Yes, for limited tasks, and no for the ideas. NIH notice NOT-OD-25-132, released 17 July 2025, applies from the 25 September 2025 receipt date. It says NIH won't consider applications "substantially developed by AI, or contain sections substantially developed by AI" to be the applicants' original ideas. The same notice says NIH "will continue to employ the latest technology in detection of AI-generated content".

The trigger was volume. NIH wrote that it had evidence of AI tools letting principal investigators submit "more than 40 distinct applications in a single application submission round". So the notice also caps each PI at six new, renewal, resubmission or revision applications per calendar year, T activity codes and R13 conference grants excepted. If AI use is found after the award, NIH may refer the case to the Office of Research Integrity and disallow costs, withhold future awards, suspend or terminate the grant. NSF moved the same way on 8 December 2025, when its policy guide began defining research misconduct as fabrication, falsification or plagiarism committed directly or through tools "including artificial intelligence (AI)-based tools".

For an R01 or an R21, split the work. You write the Specific Aims page and the Approach. The model gets the jobs that originate nothing: the check against the funding opportunity, the reviewer critique, the cut to the page limit and the plain language summary. NIH itself warns that "AI use may result in plagiarism, fabricated citations, or other kinds of research misconduct". So every reference a model touched goes back to PubMed before submission. The page on invented sources shows how those citations look when they're wrong.

How do you use AI for grant writing, from RFP to draft?

Give the model the funder's own documents first and your facts second, then run eight passes: requirements, fit, outline, need, objectives, budget, critique and compliance. Finding the funder is a different job, since a chat model isn't a grant database. Search Grants.gov for US federal money or a foundation directory for private grants, then bring the RFP back here. For a 60 page solicitation, the PDF summary method covers how to keep the model from skipping pages.

Prompt 1, the requirements matrix. Attach the RFP or funding opportunity and send: "From this document only, build a table with one row per requirement: section name, page or word limit, required headings, review criterion and its points or weight, eligibility rule, required attachments, formatting rule and deadline. Quote the line each row comes from. If something is not in the document, write not stated." Every later prompt checks against this table, so keep it in the same chat.

Prompt 2, the fit check. "Here is the requirements table, and here are our facts: [mission, service area, annual budget, past grants, staff, outcomes with numbers]. For each eligibility rule and funding priority, mark meets, partly meets or does not meet, and quote the fact you used. Do not assume capacity we did not state." A does not meet on eligibility ends the application here, before anyone writes a page.

Prompt 3, the outline. "Draft an outline that uses the funder's section headings word for word and in their order. Give each section a word budget in proportion to its review points, and under each heading list the criterion it answers and which of our facts belongs there. No prose yet." Reviewers score against the criteria, so an outline built from them is the cheapest score gain in the whole process.

Prompt 4, the statement of need. "Write the statement of need in [N] words using only the statistics below, each followed by its source in parentheses. Where a claim needs a number I did not give you, write [NEED DATA] instead of a figure." Paste the census, county health or school data yourself, with its source. The [NEED DATA] tag is what stops an invented percentage from reaching a program officer.

Prompt 5, objectives and the logic model. "Turn these program activities into three to five outcome objectives in this form: by [date], [number] of [population] will [measurable change], measured by [tool]. Then lay out a logic model as a table with inputs, activities, outputs, first year outcomes and end of grant outcomes. Use only the numbers I gave you."

Prompt 6, the budget justification. "Write a justification for each line of this budget, in the funder's cost categories: [paste lines by role and percent effort, not by name]. Say what each cost buys and why the project needs it. Do not change, round or add any figure. If a line looks hard to justify, flag it instead of inventing a reason." Check the totals against your spreadsheet afterwards. The model's arithmetic isn't the budget.

Prompt 7, the reviewer critique. "Act as a reviewer for this funder, using the criteria and points in the requirements table. Score each section, quote its weakest sentence, and say what evidence would raise the score. Do not rewrite anything." This is the pass where a model earns its place on a grant team. It keeps the ideas yours, and that's exactly the line NIH and UKRI draw.

Prompt 8, compliance and the cut. "Check this draft against the requirements table. List every missing requirement, every heading that differs from the funder's wording, and the word count of each section against its limit. Then cut [section] to [N] words without removing any number, name or citation." Run it after the last human edit, and read the cut version against the original before accepting it. The writing workflow page has the general voice and editing rules that apply to every section.

What should you never paste into an AI tool when writing a grant?

Never paste what you don't own or have no consent to share: a proposal you're reviewing, other people's personal data, and collaborators' unpublished work. The funder policies above say so in close to those words. UKRI's version is the plainest: "Sensitive or personal data of others must never be input into a generative AI tool without formal consent from the individual."

What is safe to paste: the public RFP, your organization's published facts, your own draft narrative and aggregate numbers with no one identifiable in them. Then check where the text goes. Lloyd's Register Foundation tells applicants to "Consider whether the AI tool may store, reuse or share the information you enter into it". The settings that decide that differ by app and by plan, and what never to paste into AI covers them for the main chat apps.

Which AI tool is best for grant writing?

The best AI tool for grant writing is a general chat model that reads PDFs, such as ChatGPT, Claude or Gemini, because every pass above is reading the funder's document and reworking your own text. Grant platforms such as Instrumentl and OpenGrants sell what a chat model lacks, a searchable list of funders and open opportunities. If your bottleneck is finding funders, pay for a grant database. If it's writing to a deadline, a chat subscription covers it.

Whizi is one such chat subscription, with 280+ models on one login. Its Projects fit grant work. Pin the RFP, your boilerplate and last year's report (up to 10 files of up to 10 MB each, including PDF, Word and Excel) plus up to 32,000 characters of standing instructions. Every chat in the project reads them. Starter at $15.99 a month ($10.99 billed yearly) runs ChatGPT Luna and Claude Haiku. Claude Sonnet 5 and GPT-5.6 Sol need Pro at $29.99 ($19.99 yearly). Whizi loses for a grants office that needs shared seats, because it has no team accounts. It has no funder database either. It doesn't use your prompts or files to train Whizi-owned models. Your text still reaches the provider of the model you pick, which may cache it under its own terms, so the paste rules above apply unchanged.