Grant season rewards speed, and AI grant writing looks like the obvious shortcut. A fabricated reference ruins it. Here is what NIH and NSF actually ask for, what NIH’s 2025 AI policy changed, and the one part of the process where an AI research tool belongs.
The short version
There is no “literature review” section in a federal grant application. The prior work lives inside your Significance and Approach sections, and the full citations live in a separate References Cited attachment. Reviewers score how well your project rests on that prior work.
Since September 25, 2025, NIH will not treat sections “substantially developed by AI” as your original ideas, and it names fabricated citations as a research misconduct risk. So the question is no longer “can AI write my grant?” It is “where does AI help without putting the application at risk?”
The answer for AI grant writing in 2026: use AI to find, verify, and organize the literature. Write the application yourself. Submit references you can defend.
What a federal grant application actually contains
Most AI grant writing advice skips this step, which is why so much of it tells you to generate a lit review. Here is what the two largest funders require.
NIH R01 (SF424 R&R)
| Project Summary / Abstract | 30 lines | Background, aims, methods, impact |
| Project Narrative | 3 sentences | Public health relevance in plain language |
| Specific Aims | 1 page | One central hypothesis, 2 to 4 concrete aims |
| Research Strategy (Significance, Innovation, Approach) | 12 pages | The scored core of the application |
| Bibliography & References Cited | No limit | Every citation from the plan, “relevant and current,” PMCIDs where applicable |
| Biographical Sketches | 5 pages each | Common Form plus NIH supplement |
| Budget & Justification | No limit on justification | Modular up to $250,000 per year in direct costs |
| Data Management & Sharing Plan | 2 pages | Required since January 25, 2023 |
| Facilities, Equipment, Letters of Support, Human Subjects / Animals | Per form | Environment and protections |
Standard due dates for new R01s are February 5, June 5, and October 5.
How it is scored. Under the Simplified Peer Review Framework, in effect for applications due on or after January 25, 2025, reviewers score three factors: Factor 1, Importance of the Research (Significance and Innovation), Factor 2, Rigor and Feasibility (Approach), and Factor 3, Expertise and Resources. Factors 1 and 2 get a 1 to 9 score. Factor 3 is rated sufficient or not.
Where the literature goes. The Significance instructions ask you to describe “the strengths and weaknesses in the rigor of the prior research that serves as the key support for the proposed project.” The Approach instructions ask you to justify design choices “from the scientific literature, preliminary data, or other relevant considerations.” The References Cited attachment then compiles the citations. Nowhere does NIH define a literature review section, and an unrequested attachment can violate the SF424 attachment rules.
NSF (PAPPG 24-1)
| Project Summary | 1 page | Three labeled statements: Overview, Intellectual Merit, Broader Impacts |
| Project Description | 15 pages | Must relate the work “to the present state of knowledge in the field”; separate Broader Impacts section; Results from Prior NSF Support if funded in the past 5 years |
| References Cited | No limit | Citations only, all authors in order, full bibliographic detail |
| Biographical Sketches | Via SciENcv | Mandatory format |
| Budget & Justification | 5 pages | Per organization |
| Data Management & Sharing Plan | 2 pages | Renamed in 24-1 to stress sharing |
| Mentoring Plan | 1 page | Covers graduate students and postdocs |
NSF judges two criteria, Intellectual Merit and Broader Impacts, and the first one turns on how well you place your idea against the current state of knowledge.
Why applications lose points
NIMH publishes the common reasons applications score poorly: aims that are overly ambitious or unfocused, a rationale that does not compel, too little preliminary data to establish feasibility, no discussion of pitfalls and alternatives, and careless presentation. Reviewers treat sloppy writing as a proxy for sloppy science, and the classic advice is one clear, testable hypothesis with a small number of independent aims, written for a non-specialist reviewer with about 30 minutes to spend [Bourne2006].
A reference that does not exist is the most careless presentation error there is. It is also worse than careless, which brings us to the policy.
What NIH’s 2025 policy changed for AI grant writing
On July 17, 2025, NIH released Guide Notice NOT-OD-25-132, Supporting Fairness and Originality in NIH Research Applications, effective for the September 25, 2025 receipt date onward. Three parts matter for anyone using AI in the process:
- Originality. NIH will not consider applications, or sections of applications, “substantially developed by AI” to be the applicant’s original ideas.
- Volume. Each Principal Investigator is limited to six applications per calendar year.
- Misconduct. NIH states that AI use “may result in plagiarism, fabricated citations, or other kinds of research misconduct.” If AI-developed content is detected after an award, the case may be referred to the Office of Research Integrity. Cost disallowance, suspension, or termination are all on the table.
NSF’s posture is similar. Its December 2023 notice to the research community encourages proposers to indicate AI use in the Project Description and holds them responsible for the “accuracy and authenticity” of the content, AI-assisted or not.
Why a fake reference is not a typo
Federal regulations at 42 CFR Part 93 define research misconduct as fabrication, falsification, or plagiarism “in proposing, performing, or reviewing research.” A grant application is proposing research. A 2026 analysis in Accountability in Research argues that hallucinated citations produced by generative AI can meet the misconduct bar when citations function as evidence and the author shows reckless indifference to whether they are real [Resnik2026].
How often it happens
When researchers asked ChatGPT to write literature reviews on 42 topics, it produced 636 references. Of those, 55% of the GPT-3.5 citations and 18% of the GPT-4 citations pointed to works that do not exist, and of the real ones, 43% and 24% respectively contained substantive errors [Walters2023]. Those are the odds when a general chatbot writes your background section unassisted.
Where an AI research tool belongs in a grant application
Here is the honest map. CoChat helps most in the parts reviewers score most heavily, and it stays out of the parts NIH says must be yours.
| Significance (Factor 1 / Intellectual Merit) | Yes | Deep Research across multiple scholarly databases maps the state of the field. A literature review table sorts prior work by what each study showed and where its rigor is weak, which is exactly what the Significance instructions ask you to describe. |
| Innovation (Factor 1) | Partial | The same map shows what has and has not been tried, so your innovation claim is defensible. The claim is yours. |
| Approach (Factor 2) | Partial | Verified sources to justify design, sample size, and controls from the literature. Preliminary data comes from your lab. |
| References Cited | Yes | Every source is checked against CrossRef as it enters your project, with a real DOI. Export to BibTeX, RIS, or CSV, or sync to Zotero. You add PMCIDs where NIH requires them. |
| Specific Aims, Project Summary, Narrative | No, by design | Under NOT-OD-25-132 the narrative must be your original work. CoChat does not draft it. |
| Biosketch, Budget, Facilities, Letters, DMS Plan, Mentoring Plan | No | Administrative and institutional. Your grants office is the right partner here. |
| Results from Prior NSF Support | Partial | Your Library holds your own prior publications. The summary is yours. |
The one-sentence rule for AI grant writing: research with CoChat, write it yourself, cite only what is real.
An AI grant writing workflow that respects the rules
1. Map the field before you write a word
Open a Project in CoChat for the application and run Deep Research on your central question. CoChat searches real scholarly databases, returns papers with verifiable DOIs, and reads full text where an open-access copy exists. Ask it to organize the results into a literature review table with columns for design, sample, key finding, and limitations. That table is the raw material for the “strengths and weaknesses in the rigor of prior research” paragraph NIH asks for.
2. Verify every reference on the way in
Each paper added to the project is checked against CrossRef, and the verification status is visible on every entry. If a source cannot be resolved, it does not silently become a citation. This is the safeguard NIH itself named when it flagged fabricated citations, and it is the step a general chatbot skips. If you want the longer version of how verification works, see How to Find Sources with AI (and Verify Every Citation).
3. Keep the field current until submission
A Research Question in the Library keeps screening new papers on your topic, so a study published the week before the deadline does not become the reviewer’s “the applicants appear unaware of…” comment.
4. Export your References Cited
Send the verified set to BibTeX, RIS, or CSV, or sync to Zotero and format from there. Check the funder’s citation format requirements and add PMCIDs where NIH requires them.
5. Write the narrative yourself
Specific Aims, Significance, Innovation, Approach: these are your ideas in your words. That is the requirement, and it is also the point. Reviewers are scoring your thinking. Use the verified table beside you, not a generated draft in front of you.
6. Disclose where the funder asks
For NSF, indicate in the Project Description how AI was used. For NIH, keep a record of your process. If your institution has an AI disclosure policy, follow it. A one-line disclosure that AI tools were used for literature search and citation verification, with all text written by the applicants, is accurate and easy to defend.
What CoChat will not do for your grant
Faculty readers deserve the plain list. CoChat will not write your Specific Aims, draft your Significance section, build your budget, produce your biosketch, or write your Data Management and Sharing Plan. It will not add PMCIDs automatically. It finds, verifies, and organizes the literature so that every reference under your argument is real and current, and it leaves the argument to you.
That division is what responsible AI grant writing looks like. It is what the 2025 policy requires, and it is how strong applications were written before AI arrived.
Before October 5
If you are in the current NIH cycle, the highest-return hour you can spend this week is a verification pass on your existing References Cited. Import the list into a CoChat project, let CrossRef check each entry, and fix what fails. Then spend the remaining days on the science.
Verify your references in CoChat
Sources
Official policy and instructions
- NIH Guide Notice NOT-OD-25-132, Supporting Fairness and Originality in NIH Research Applications (released July 17, 2025). https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-132.html
- NIH Guide Notice NOT-OD-24-010, Simplified Peer Review Framework (effective for due dates on or after January 25, 2025).
- NIH SF424 (R&R) Application Guide, Research Strategy (G.400) and Bibliography & References Cited (G.220) instructions.
- NIH Guide Notice NOT-OD-21-013, Data Management and Sharing Policy (effective January 25, 2023).
- NIH Standard Due Dates for Competing Applications.
- NIMH, Common Mistakes in Writing Applications.
- NSF Proposal & Award Policies & Procedures Guide (PAPPG), NSF 24-1, Chapters II.D.2 and III.A.2.
- NSF, Notice to the Research Community on the Use of Generative Artificial Intelligence (December 14, 2023). https://www.nsf.gov/policies/ai/merit-review
- 42 CFR Part 93, Public Health Service Policies on Research Misconduct.
Peer-reviewed
- Walters WH, & Wilder EI (2023). “Fabrication and errors in the bibliographic citations generated by ChatGPT”. Scientific Reports. https://doi.org/10.1038/s41598-023-41032-5
- Resnik DB, & Hosseini M (2026). “Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papers”. Accountability in Research. https://doi.org/10.1080/08989621.2026.2645390
- Bourne PE, & Chalupa LM (2006). “Ten Simple Rules for Getting Grants”. PLoS Computational Biology. https://doi.org/10.1371/journal.pcbi.0020012
Every source in this article was checked against CrossRef or read at the primary source before publication.
