An AI draft reads smoothly, and that is the risk. A made-up statistic, a program officer who doesn't exist or a citation to a report nobody wrote all look exactly like the true sentences around them. If one reaches a reviewer, your organization's credibility goes with it.
This article gives you a fact-checking routine for any draft that started with a tool: what tends to go wrong, how to trace each claim, the order to check things in and what to do with whatever you can't confirm.
Why drafts contain false statements
NIST's Generative AI Profile uses the term "confabulation" for a phenomenon in which generative AI systems generate and confidently present erroneous or false content in response to prompts. NIST notes that these systems are designed to predict likely words, which can produce accurate output but can also produce inaccurate or internally inconsistent output. NIST calls this particularly relevant for open-ended prompts that ask for long-form responses, which describes a grant narrative.
NIH makes the same point in its originality notice, NOT-OD-25-132: it warns that AI use can lead to plagiarism and fabricated citations. So the risk is not limited to one product or one kind of error. Plan for it in every draft.
Grant drafts tend to go wrong in a few predictable places:
- Statistics with no source, or a real source quoted with the wrong number
- Citations to studies, laws or reports that don't exist or don't say what the draft claims
- Names of partners, staff, program officers or board members
- Dates and counts, such as the year you were founded or the number of people served
- Claims about the funder's priorities that don't appear in the notice
You answer for the text
Funders hold the applicant responsible, whatever tool helped. NSF states that proposers are responsible for the accuracy and authenticity of their proposal submission, including content developed with generative AI tools. And when subrecipients apply for funds, 2 CFR 200.415(b) has them certify to the pass-through entity that the information is true, complete and accurate, and acknowledge that false information or the omission of a material fact can bring criminal, civil or administrative consequences. Whoever signs is vouching for the text.
Build a claim log
Before editing for style, read the draft and pull out every checkable claim. A spreadsheet or table is enough. For each one, record the claim, where it came from and who verified it.
Say a county health nonprofit gets a draft sentence: "In Harlow County, 34% of households with children live below the poverty line, according to the Census Bureau." Here is how the log entry might look after checking:
| Claim | Source in hand | Checked by | Result |
|---|---|---|---|
| 34% of households with children below poverty | None attached to the draft | Program director | Census table shows 29% for the county; correct to 29% and cite the table |
| Served 410 families in the program year | Annual report, p. 6 | Development director | Report says 410; keep |
| Founded in 2009 | Board bylaws | Executive director | Bylaws say 2011; correct |
Three claims, one confirmed and two corrected. The draft would have submitted two errors with no sign that anything was wrong.
The rule behind the table is simple: every claim must trace to a document you can open and show someone. If you can't find the document, the claim isn't ready.
Check in this order
Work through the draft in four passes, from the most checkable to the least.
1. Numbers
Compare each figure with your own records first: the annual report, program data, the audit, the budget. For community statistics, go to the original table rather than a summary. If you pull data from the Census Bureau's American Community Survey, note which estimate you used. The Census Bureau lists 1-year estimates for areas of 65,000 people or more and 5-year estimates for all areas, so a small county or a census tract will need the 5-year data. Cite the table and the year range in the text.
2. Names
Check that every person, partner and agency exists and is spelled and titled correctly. Match staff names to your own staff list and partners to signed letters or agreements. For another nonprofit, check its Form 990, which tax-exempt organizations must make available for public inspection. Look up officials on their organization's own site.
3. Dates
Founding dates, award dates and project timelines should match a document. Check that the project calendar fits the period of performance in the notice, not only the draft.
4. Citations
Open every cited source. Confirm that it exists, that the number or quote appears in it and that it says what the draft claims. If the draft names a law or regulation, read the text in the official source. If you can't find the source in a few minutes, remove the citation and the claim that depends on it.
Handle what you can't confirm
Each unconfirmed claim has three possible outcomes:
- Replace it with a verified fact from your records or an official source.
- Soften it to something you can support, such as "many of the families we serve" in place of a percentage.
- Cut it. A proposal with a missing statistic is better than one with an invented one.
Never leave a placeholder like "[statistic]" or "[cite source]" in a document that is about to go out. Search the final file for brackets before you submit.
Keep the files that back each claim
Record where each verified fact came from, in the log or in a folder for the application. When a program officer asks about a number during review or after an award, you can answer in a minute. Material that you already maintain helps here too: a boilerplate library of approved figures gives you one place to check against, and the final read-through is the right moment for one last pass over the claim log.
Fact-checking is also a place to share the work. A program director can verify program numbers and a finance lead can verify the budget, so you aren't the only one checking every line.
Your next step
Take the next draft you have, AI-assisted or not, and build a claim log for the first two pages. Verify each claim against a document you own or an official source, and note how many needed a correction. That count tells you how much checking your drafts need.
A tool that drafts from your own files makes this routine easier, because you know where claims should come from. What to hand off to AI and what to keep explains why the checking stays with you, and keeping your organization's voice covers how to feed a tool your own material.
SoliContraX drafts from your organization's files and flags names, contacts and citations they don't support, so you know where to look first. The checking and the final call stay with you. Start your next application.
Sources
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, section 2.2 Confabulation (NIST)
- NOT-OD-25-132, Supporting Fairness and Originality in NIH Research Applications (NIH)
- Notice to Research Community: Use of Generative Artificial Intelligence Technology in the NSF Merit Review Process (NSF)
- 2 CFR 200.415, Required certifications (eCFR)
- Which ACS estimates to use (U.S. Census Bureau)
- Public disclosure overview (IRS)