Using ChatGPT to Review Grant Proposals Against a Rubric
A step-by-step way to check a grant draft against the funder's rubric with ChatGPT, where a model's scores mislead, and what funders tell their own reviewers about AI.

In this article (9)
- Quick Answer: How Do You Review a Grant Proposal With ChatGPT?
- How to Review a Grant Proposal With ChatGPT, Step by Step
- Reading the Model's Output: A Worked Example
- What to Ask ChatGPT to Check in Each Section
- Where ChatGPT Rubric Scores Go Wrong
- What Funders Tell Their Own Reviewers About AI
- AI Proposal Review Tools Built Into Grant Software
- FAQ
- Pair the Model's Check With a Human Read
Using ChatGPT to review grant proposals against the funder's rubric means giving the model the announcement's review criteria and your draft, then asking where each criterion is answered and where the answer is thin. One commenter on Reddit described the routine this way: "I upload the RFA, scoring rubric, and other info and ask it to review and score the proposal" (Reddit thread 1ubd86x, read September 2026).
The method is useful within limits. A model can find the criterion your draft never answers. It cannot tell you how a panel will score you, and one commenter in the same thread reported "major discrepancies between LLM scores and community reviewers and expert reviewers" (Reddit thread 1ubd86x, read September 2026).
Quick Answer: How Do You Review a Grant Proposal With ChatGPT?
Give ChatGPT the funder's review criteria and a draft stripped of confidential details, then ask it to quote the passage that answers each criterion and name the gaps. Treat its scores as a checklist aid, not a prediction: the model is not your review panel, and NIH bars its own peer reviewers from using generative AI to write critiques.
Writing your own? Start it as a draft in GrantCue, with AI help for each section. Start the proposal free
How to Review a Grant Proposal With ChatGPT, Step by Step
A rubric review with ChatGPT works best as an evidence search: the model finds the passage that answers each criterion, and you decide whether the answer is strong enough. It checks a draft you have already written; our guide on how to write a grant proposal covers writing each section.
Start from the funder's own criteria, copied word for word with any point values, in the order the announcement lists them. Our guide to how reviewers score proposals against a rubric explains where those criteria sit and how scales differ by agency. When a funder lists questions or priorities without point values, use its application questions as the criteria, in its order, and weight them equally until the funder says otherwise. Then run the review in steps, one request at a time.
Asking for quotations is the control. A model that has to quote your draft can be checked in seconds, by searching the draft for the quoted words. A model that only offers an opinion cannot be checked at all.
The commenter who uploads the RFA and rubric described the aim: "I use ChatGPT as a grant reviewer to review proposals and look for weaknesses before submitting," adding "I don't upload any confidential or sensitive information" (Reddit thread 1ubd86x, read September 2026). Another commenter in the thread pushed back: "You could be violating privacy by uploading it to ChatGPT unless it's a non-training version" (Reddit thread 1ubd86x, read September 2026).
The rubric from a public announcement carries no confidentiality risk. Your draft can. On OpenAI's services for individuals, "we may use your content to train our models" (OpenAI Help Center, read September 2026), so strip the draft before it goes into a personal account.
Reading the Model's Output: A Worked Example
The output is worth only what your checks make of it, so read each row as a claim to verify. A composite review makes the point: one draft checked against a five-criterion, 100-point rubric, with what the model returned and what the writer found on checking (illustrative figures).
Illustrative figures, not data from a real organization.
Two rows show why the checks matter more than the model. The capacity quotation was a paraphrase the model presented as your text, and the approach answer existed all along in an attachment the model never saw. The most useful row is the empty one: in this composite rubric, 35 of the 100 points ride on a section the narrative barely covers (illustrative figures).
Log each row as you work it. A table with the criterion, the passage, your verdict and the next action becomes the edit list for the next draft, and it keeps the model's opinions out of the proposal itself.

What to Ask ChatGPT to Check in Each Section
Run the check section by section, since each section answers different criteria and a single prompt for the whole proposal blurs them together. Three sections repay the most attention.
Reviewing a Grant Proposal Executive Summary Using AI
An executive summary is read first and shapes how a reviewer reads everything after it. Ask the model which terms from the rubric appear in your summary and which are missing, and whether the summary states the need, the approach and the amount requested. Then rewrite the summary yourself, in the funder's terms and your own sentences. Our executive summary examples show the structure a reviewer expects.
Checking an Evaluation Plan, Including CBPR Principles
For an evaluation plan, ask the model to build a table from your text: each outcome, its measure, the data source, who collects the data, when, and the target. Empty cells are your gaps. If the funder asks for an evaluation grounded in community-based participatory research (CBPR) principles, paste the funder's own description of those principles into the prompt and ask where the plan answers each one, because the model's general idea of CBPR is not the funder's. The evaluation plan examples show a complete plan to compare against.
Testing the Method on a Sample Grant Proposal First
Before you run your own draft, run a public sample proposal through the same prompts. Read the sample yourself first and note its gaps, then see whether the model finds the same ones or invents new ones. That test gives you a first read on how far to trust the model with your real draft. Our sample grant proposals for nonprofits work well for this.
Where ChatGPT Rubric Scores Go Wrong
ChatGPT will produce a score if you ask for one, but nothing ties that number to how a panel will score the same draft. The number is a prompt to reread, not a forecast. A quick test shows how much weight it can bear: run the same scoring prompt twice on the same draft and compare the results. If the scores move between runs, treat the number as noise and keep the quotations and gaps.
One commenter in a Reddit thread said "I'm skeptical about LLM's ability to score grants," citing "major discrepancies between LLM scores and community reviewers and expert reviewers" and a declined grant that "received strong LLM scores but mixed community review scores" (Reddit thread 1ubd86x, read September 2026). That is one person's experience, not a measured error rate. It still fits how the tool works: a model reads only what you paste, and it knows nothing of the panel, the round or the funder's unwritten priorities.
Source for the last row: OpenAI Help Center, educator guidance on AI-generated content (read September 2026).
Watch the model's suggestions as closely as its scores. GrantStation, a grant-database vendor, warns that AI "can invent statistics" (GrantStation, Grantwriting with AI Starter Kit, read September 2026), and a review prompt that asks for stronger evidence invites exactly that. Any figure the model proposes needs a source you found yourself.
Watch its quotations too. A model asked to find a passage can return a paraphrase that sounds like your draft but is not in it. Search for every quotation before you accept that a criterion is covered.

What Funders Tell Their Own Reviewers About AI
NIH, NSF and NEH restrict how their peer reviewers use generative AI, and UKRI and the Wenner-Gren Foundation tell reviewers not to guess whether an applicant used it. These rules bind reviewers, not applicants, and a model's critique is not a rehearsal of the one a panel will write.
NIH's notice NOT-OD-23-149 "prohibits NIH scientific peer reviewers from using natural language processors, large language models, or other generative Artificial Intelligence (AI) technologies for analyzing and formulating peer review critiques for grant applications and R&D contract proposals," and extends the rule to members of NIH National Advisory Councils and Boards (NIH notice NOT-OD-23-149, released June 23, 2023). NSF states: "NSF reviewers are prohibited from uploading any content from proposals, review information and related records to non-approved generative AI tools" (NSF notice, December 14, 2023). NEH "prohibits peer reviewers from uploading NEH applications into a cloud database" (NEH AI policy, October 23, 2024). The Community Foundation of Henderson County says its volunteer reviewers "do not use AI in their evaluation process" and "find there is no substitute for personal and authentic responses and perspectives in applications" (Community Foundation of Henderson County, AI Usage Statement, modified May 20, 2026).
Some funders also tell reviewers not to guess at AI use. UKRI assessors must "not take into account or speculate within their assessment whether generative AI has been used to develop the application" (UKRI generative AI policy, last updated July 6, 2026). The Wenner-Gren Foundation says "reviewers should not try to determine if Generative AI was used by applicants but evaluate all proposals on their own merits" (Wenner-Gren Foundation AI policy, March 19, 2025).
AI does appear on the funder's side, on the funder's terms. Seven UK funders say they "may use enterprise data protected (EDP) generative AI tools to process funding applications" (UK funders' joint statement, revised June 2026). Submittable's Smart Reviewer, a funder-side tool that "evaluates applications based on the criteria you set in your Custom Feedback Form," is described as "currently in Beta," and "Only organizations that have signed the Beta Agreement can have this feature turned on" (Submittable support, updated July 15, 2026).
Either way, the published criteria are the one part of the review you can see in advance. That is why the rubric, not the model's opinion of your prose, belongs at the center of your own check.

AI Proposal Review Tools Built Into Grant Software
Some grant products build AI review into the software, and their own framing treats the output as support, not a verdict.
GrantStation members can "upload a Word document or PDF to the Proposal Reviewer for AI-assisted feedback on strengths, areas for improvement," under the stated stance "Use AI as support, not a substitute" (GrantStation member benefits page, read September 2026).
GrantCue, which publishes this guide, sells one of these tools. Its AI Grant Writing Workspace includes "Funder alignment analysis and scoring," and its AI scoring page describes an Opportunity Fit and Application Readiness pair that "Separates whether an opportunity aligns with your organization from whether the application is ready to pursue" (grantcue.com, September 2026). The same page states: "Neither score predicts whether you will win." Each fit and readiness score shows "Evidence-coverage confidence," which the page says "is not a statistical confidence interval," along with the exact sources used and "every material gap such as an unlinked Form 990 or missing draft" (grantcue.com, September 2026). The AI scoring explainer sets out how each score is built.
Whatever the tool, the useful output is the same: the criteria, the passage that answers each one, and the gaps that remain. A tool that shows its evidence can be checked the same way as a chat model, by finding each cited passage in your draft.
FAQ
Can ChatGPT Score a Grant Proposal?
It will produce a score if you ask, but the number is not a forecast of your panel's score. One commenter on Reddit reported "major discrepancies between LLM scores and community reviewers and expert reviewers" (Reddit thread 1ubd86x, read September 2026). Use the score to decide which criteria to reread, and base your edits on the quoted passages and the listed gaps.
Do NIH Reviewers Use ChatGPT?
NIH prohibits it. Notice NOT-OD-23-149 bars NIH scientific peer reviewers from using large language models or other generative AI "for analyzing and formulating peer review critiques for grant applications and R&D contract proposals," and extends the rule to NIH National Advisory Councils and Boards (NIH notice NOT-OD-23-149, released June 23, 2023). The rule covers reviewers; applicants follow separate NIH guidance.
Should I Upload the Scoring Rubric to ChatGPT?
A rubric copied from a public announcement carries no confidentiality risk, so include it word for word with any point values. The draft is the part to protect. On OpenAI's services for individuals, "we may use your content to train our models" (OpenAI Help Center, read September 2026), so remove names, salaries and unpublished data first, or use an account whose terms you have checked.
Can ChatGPT Tell If a Grant Proposal Was Written by AI?
No. OpenAI says ChatGPT's answers to questions like "could this have been written by AI?" are "random and have no basis in fact" (OpenAI Help Center, read September 2026). OpenAI also withdrew its own AI classifier "due to its low rate of accuracy" (OpenAI, classifier retirement note dated July 20, 2023). Leave that question out of a rubric review and spend your prompts on the criteria.
What Should I Do With ChatGPT's List of Weaknesses?
Sort it. Weaknesses tied to a named criterion and a quoted passage are worth checking against the rubric; general style advice can wait. Verify each claimed gap in the draft itself, since the answer may sit in an attachment the model never saw. Any statistic it suggests adding needs a source you find yourself, because GrantStation warns AI "can invent statistics" (GrantStation, read September 2026).
Pair the Model's Check With a Human Read
Hand the criteria table and the draft to a colleague who has not seen the proposal, and ask for what you asked the model: the passage that answers each criterion, and the gaps. Where the two reads disagree, test the colleague's point first.
After a decision, the funder's own comments are the best rubric feedback you will get. Our guide to asking a funder why an application was declined covers how to request them. File the criteria table with the proposal, dated and marked with the draft version it checked, so the next round starts from it.