Chapter 8: Grant Writing with AI
What You Will Learn
How grant writing is strategically different from other research writing, and why that changes how you should use AI.
Which parts of a grant proposal AI genuinely helps with, and which parts it cannot do for you.
What “substantially developed by AI” actually means under NIH policy, and how to stay on the right side of that line.
What University of Michigan, NIH, and NSF policies currently say about AI use in proposals.
A “Try This” exercise at the end to practice using AI as a strategic thought partner on your own proposal.
Why Grant Writing Is Different
If you have already read the chapter on AI-assisted writing, you might wonder whether this chapter is redundant. The short answer is no, and the reason comes down to what grant writing actually is.
Writing a manuscript is primarily an act of communication. You have findings. Your job is to describe them clearly, accurately, and in a way that helps readers understand their significance. AI can genuinely help with that: organizing your argument, smoothing your prose, adapting your language for a particular journal audience.
Grant writing is something else. It is a strategic act. You are not reporting what you found — you are making an argument for why a specific research question matters, why your team is the right one to answer it, and why the approach you have chosen is sound. The strength of a grant proposal comes directly from the depth of your thinking about those three things. A beautifully written proposal built on shallow thinking will not survive review. A proposal with clunky prose but a genuinely original and well-argued question often will.
This means AI’s role in grant writing is less about the writing and more about the thinking. The rest of this chapter is organized around that distinction.
Where AI Actually Helps in Grant Development
Mapping the Landscape Before You Write
One of the most useful things AI can do in the early stages of grant development is help you get oriented quickly. If you are applying to a new funding mechanism, responding to a specific program announcement, or writing for an audience outside your primary discipline, you can use a tool like UM-GPT or ChatGPT to summarize what a particular study section typically values, what common weaknesses reviewers flag in your target mechanism, or how similar funded projects have framed their significance. This is not cheating — it is using AI the same way you would use a mentor’s advice or a funded application from your institution’s grants office.
Similarly, AI can help you scan and organize a large literature quickly during the aims development stage. If you are trying to identify the specific gap your work addresses, you can describe your area to an AI tool and ask it to highlight what remains understudied. You will still need to verify everything it surfaces, and you will likely find that it misses things a field expert would catch immediately. But it can give you a working map to react to, which is often faster than starting from scratch.
Stress-Testing Your Aims
This is one of the most underused applications of AI in grant writing, and it is genuinely valuable. Once you have a draft of your specific aims, you can ask an AI tool to take the role of a skeptical reviewer and identify the weaknesses. What would a study section worry about? What assumptions in your approach are most vulnerable? What is the most obvious alternative explanation for the expected findings?
The AI’s critique will not be as sharp as a real grant reviewer’s, and it will miss things that only someone in your subfield would catch. But it will often surface the generic structural problems that are easy to miss when you have been staring at the same document for weeks. Think of it as a first-pass review before you ask colleagues to read it.
Timelines, Budgets, and Boilerplate
Gantt charts, milestone tables, resource justifications, facilities descriptions — these are real parts of a grant proposal and they take time. AI tools can help draft these efficiently. There is very little intellectual risk here. The substance of your timeline and budget still has to reflect your actual project, but AI can give you a reasonable starting structure to fill in rather than formatting a table from scratch.
Language Polishing
This is where the overlap with the writing chapter is most direct. AI can help with clarity, coherence, flow, and meeting page limits. If you have a 12-page research strategy that needs to fit into 6 pages, AI can help you identify what to cut and how to compress without losing key content. If English is not your first language, AI can help make your prose more natural without changing your scientific meaning. These are legitimate uses, and they are broadly accepted.
What AI Cannot Do
AI cannot provide the intellectual substance of your proposal. It cannot tell you what the right research question is, whether your experimental design is actually sound, or why your preliminary data supports the approach you are proposing. It cannot make the tacit connections that come from years of working in a field — the sense of which questions are genuinely open versus technically open, which methods are reliable in your specific context versus in general, which claims will ring true to your study section and which will raise flags.
More importantly: if you cannot walk into a room of experts in your field and defend every aim, every methodological choice, and every framing decision without the AI to help you, then the proposal does not yet reflect your intellectual contribution. That is the standard that matters, both ethically and practically.
What “Substantially Developed by AI” Actually Means
In 2025, NIH issued Notice NOT-OD-25-132, which states that applications “substantially developed by AI” may be considered non-original and subject to rejection [National Institutes of Health, 2025]. The policy also limits PIs to six applications per year starting September 2025. Understanding what triggered this policy helps explain what the line actually is.
The immediate context was a researcher who submitted approximately 40 NIH applications in a single grant cycle — a volume that is essentially impossible through normal human writing. The sheer number made it obvious that AI was doing the generative work, not just assisting with it. That volume also flooded the review system in a way that harmed other applicants and degraded the peer review process for everyone.
But the policy is not just about volume. The deeper issue is about whose intellectual contribution a proposal represents. NIH funds you — a specific person or team with specific expertise — to answer a specific question. The proposal is supposed to be the evidence that you have thought deeply enough about the question to be trusted with the funding to pursue it.
Where Is the Line?
Think of it in two layers. The first layer is the scientific core: the research question, the central hypothesis, the experimental design, the interpretation of what makes your work significant and innovative. If AI generated these elements, the application misrepresents whose intellectual contribution it reflects, regardless of how well-written it is.
The second layer is expression: sentence structure, flow, clarity, and how you communicate ideas you already have. AI assistance here is much closer to what a writing center consultant, a mentor who reads your drafts, or a grammar tool does. That kind of assistance has always been part of the grant writing process.
The practical test: if you could sit in front of a study section and defend every aim, every methodological choice, and every framing decision without the AI in the room, the proposal reflects your intellectual contribution. If you could not, it does not.
NOT-OD-25-132 does not ban all AI assistance outright. The notice states that AI tools “may be appropriate to assist in application preparation for limited aspects or in specific circumstances” [National Institutes of Health, 2025]. What it prohibits is proposals substantially developed by AI, which it treats as non-original work. Critically, the policy has teeth beyond submission: if AI use is detected after an award is made, NIH may refer the matter to the Office of Research Integrity for a misconduct investigation and may revoke funding [National Institutes of Health, 2025].
Institutional and Funder Policies
If You’re at U-M
U-M guidance emphasizes using institutionally governed tools rather than public AI platforms when developing proposals, not uploading sensitive or unpublished research material to commercial systems, and documenting AI use when required by funders or journals [University of Michigan, 2025, of Michigan, 2025]. In practice, this means working within UM-GPT or a similarly governed environment when handling preliminary data, unpublished findings, or confidential collaborator information. See AI Resources at the University of Michigan for a full list of approved platforms.
NIH
Under NOT-OD-25-132 [National Institutes of Health, 2025], applications that are substantially developed by AI are not considered original work and will not be reviewed. AI tools may assist with limited aspects of preparation, but applicants must certify that the application represents their own original ideas. The stakes extend beyond rejection: if AI-generated content is detected after an award is made, NIH may refer the matter to the Office of Research Integrity for a misconduct investigation and revoke funding. Separately, NIH prohibits AI use in peer review under NOT-OD-23-149 [National Institutes of Health, 2023] — this applies to reviewers, not applicants, but is worth knowing.
One important caveat: NIH policy in this area is evolving quickly. The notice issued in 2025 was partly a response to an emerging situation, and further guidance is likely. Always check the current notice before submitting.
NSF
NSF’s position, articulated in a 2023 community notice [National Science Foundation, 2023], permits proposers to use generative AI but places full responsibility for accuracy and authenticity on the proposer. Confidential proposal content cannot be uploaded to commercial AI tools. Reviewers cannot use AI for proposal evaluation. Fabrication, falsification, or plagiarism remains research misconduct under PAPPG regardless of whether AI was involved.
Other Agencies
A systematic review of federal grant notices published between 2009 and 2024 found that most agencies have set very few explicit AI requirements or restrictions in their grant programs [Bateyko and Levy, 2025]. For grant applicants specifically, the picture as of early 2026 is similar. DOE has published extensive internal AI governance guidance — including a Generative AI Policy issued in December 2025 — but that guidance is directed at agency employees and contractors, not at grant applicants [U.S. Department of Energy, Office of the Chief Information Officer, 2025]. Its focus is on internal compliance and responsible use within DOE operations, not on proposal authorship integrity. DOD and DARPA have not issued specific guidance on AI use in proposal writing; their policy attention is on research security and foreign influence risks. NASA and NOAA have not issued applicant-facing AI disclosure requirements as of this writing.
In the absence of specific guidance, the safe approach is to follow U-M’s institutional policies and apply general research integrity principles: transparency, originality, and protection of confidential information. This area is evolving quickly, and agency-specific requirements may have been issued since this chapter was last reviewed. Always check the current program announcement or solicitation before submitting.
Using AI Responsibly in Grant Writing
The earlier sections cover this in pieces, so here is a quick synthesis. AI is appropriate for orientation work — mapping the literature, understanding a funding mechanism — and for structural feedback like stress-testing your aims. It is also appropriate for administrative writing (timelines, facilities descriptions) and for polishing language around ideas you already have. What it is not appropriate for is generating the research question, central hypothesis, or experimental rationale. If those came from the AI rather than from your own thinking, the proposal does not reflect your intellectual contribution, regardless of how well it reads.
On the practical side: keep your work in UM-GPT or another institutionally governed environment when handling preliminary data or unpublished material. Document what you used at each stage. And always check your funder’s current submission guidelines before submitting — the policies in this chapter were accurate at the time of review, but this is a fast-moving area.
A Note on Transparency
Whatever tools you use, document them. Keep a record of which AI tools you used, at what stages, and for what purposes. Some funders are beginning to require disclosure, and even where disclosure is not yet required, having that documentation protects you if questions arise later. The AI Usage Cards framework provides a structured way to do this. Chapter 11 covers how to maintain a Usage Card throughout your project and how to translate it into a methods disclosure at submission (see Checking AI Output).
Transparency also means being honest with yourself about where the intellectual work is coming from. If you find yourself accepting AI-generated aims without being sure why they are structured the way they are, that is a signal to step back and do more of the thinking yourself before moving forward.
Try This
The following exercise works best if you have a grant idea you are actually developing, but it also works with a hypothetical project.
Step 1. Write a one-paragraph description of your research question, why it matters, and what you plan to do. Write it yourself, without AI, in whatever rough form comes naturally.
Step 2. Share that paragraph with UM-GPT or another AI tool and ask: “You are a skeptical NIH study section reviewer. What are the three most significant weaknesses in this aims concept?” Read the response carefully. Note which criticisms feel valid and which feel off-base.
Step 3. Now ask the AI to draft a two-sentence version of your significance statement. Compare it to your own. Where is the AI’s version stronger? Where has it lost something important that only you would know to include?
Step 4. Revise your paragraph based on your own reactions to both responses. The goal is not to use the AI’s language — it is to use your reaction to the AI’s language to sharpen your own thinking.
This exercise illustrates the role AI works best in: not as a generator of your scientific ideas, but as a structured sounding board that pushes you to articulate and defend them more clearly.
Last reviewed: April 2026. Funder policies cited in this chapter — particularly those from NIH and NSF — are subject to change. If you notice outdated content, open an issue on GitHub.
References
National Institutes of Health. Supporting fairness and originality in nih research applications. Notice NOT-OD-25-132, 2025. Accessed 2025-12-08. URL: https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-132.html.
Office of the Vice President for Research University of Michigan. Ovpr guidelines for the use of generative ai in internal peer review. 2025. Accessed 2025-12-08. URL: https://research.umich.edu/wp-content/uploads/2025/08/OVPR-AI-Guidelines-Internal-Peer-Review_2025.pdf.
University of Michigan. Generative ai resources. UM-GPT Resource Website, 2025. Accessed 2025-12-08. URL: https://genai.umich.edu/resources.
National Institutes of Health. The use of generative artificial intelligence technologies is prohibited for nih peer review. Notice NOT-OD-23-149, 2023. Accessed 2025-12-08. URL: https://grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html.
National Science Foundation. Notice to the research community on use of generative artificial intelligence. Policy Statement, 2023. Accessed 2025-12-08. URL: https://www.nsf.gov/news/notice-to-the-research-community-on-ai.
Dan Bateyko and Karen Levy. One bad nofo? ai governance in federal grantmaking. arXiv preprint, 2025. Accessed 2025-12-08. URL: https://arxiv.org/abs/2505.08133.
U.S. Department of Energy, Office of the Chief Information Officer. Department of energy generative artificial intelligence policy. December 2025. Accessed 2026-04-06. URL: https://www.energy.gov/cio/articles/department-energy-generative-artificial-intelligence-policy.