AI can draft a grant narrative or a donor letter in minutes. It can also fill that draft with numbers, studies, and funder details that sound exactly right and are not real. Here is why that happens, what it can cost a nonprofit, and the habit that catches it.
Educational guide · written for nonprofit staff and board members · current as of September 2026 · not legal, tax, or compliance advice
Why AI invents facts
A general-purpose chatbot does not look anything up unless it has been connected to a search tool or given documents and told to use them. On its own, it writes by predicting which words are likely to come next. A statistic ("one in five families..."), a study citation, a foundation's funding priorities, or a submission deadline all follow familiar patterns, so the model can produce something that fits the pattern whether or not it matches reality.
OpenAI's own researchers published an explanation of this in September 2025. Their argument is that common training and evaluation methods reward a model for guessing rather than for admitting it does not know. The practical upshot for you: a model delivers an invented figure in the same confident tone as a correct one, and the tone tells you nothing.
The mental model
Treat a general-purpose AI like a fast, articulate volunteer who has read a great deal, never says "I'm not sure," and never checks a source. Helpful for a first draft. Never the source of a fact.
What fabrication looks like in nonprofit work
The risky errors are rarely obvious. They tend to look like this:
Invented statistics. A precise-sounding figure about need, prevalence, or outcomes, with no real source behind it.
Phantom research. A study, report, or journal article that does not exist, or a real one described as finding something it does not.
Misattributed numbers. A real figure credited to the wrong organization, year, geography, or population.
Invented funder details. A foundation's "stated priorities," typical award size, or program officer name that the funder never published.
Wrong eligibility rules and deadlines. A plausible but incorrect eligibility requirement, page limit, match requirement, or due date.
Stale facts. A figure that was accurate years ago and has since been revised or superseded.
Illustrative examples (hypothetical, not real incidents)
Grant proposal. A development associate asks AI to strengthen a statement of need. The draft adds a county-level food insecurity rate attributed to a named national study. The study exists, but it does not report county figures, and the number appears nowhere in it.
Impact report. A program manager asks AI to "summarize our outcomes" from rough notes. The summary rounds 61 participants up to "more than 100 families served" and adds a retention percentage that no one ever measured.
Donor appeal. A communications lead asks AI for a year-end letter. It includes a quotation "from a recent participant" that no participant said.
Each of these reads smoothly and would survive a quick skim. None survives a check against the source.
What it can cost
For a nonprofit, credibility is the asset that makes every other relationship work. A funder, reviewer, or journalist who finds one invented figure has reason to doubt the rest of the document, and often the organization.
Documented cases outside the nonprofit sector show how quickly this surfaces. In 2025, Deloitte agreed to refund part of its fee to Australia's Department of Employment and Workplace Relations after a report it prepared was found to contain references to research that does not exist and a fabricated quotation from a court judgment; a revised version disclosed that a generative AI tool had been used. Also in 2025, the U.S. "Make America Healthy Again" report was found to cite studies that do not exist, and the published version was later revised to replace them. In both cases outside readers, not the authors, caught the errors.
Funders are responding. The National Institutes of Health issued a notice in July 2025 stating that it will not treat applications substantially developed by AI as the applicant's original work, naming fabricated citations as one of the risks, and listing possible post-award consequences that include disallowed costs and termination of the award.
Accuracy obligations also run through grant reporting. For federal awards, the Uniform Guidance requires financial reports to carry a signed certification that the report is true, complete, and accurate, and warns that false information may carry criminal, civil, or administrative penalties. Private grant agreements commonly include their own reporting requirements. Whether a specific error creates legal exposure depends on the facts and the agreement, so confirm with counsel if a problem surfaces in something already submitted.
It is not only proposals
Invented facts turn up in board decks, annual reports, website copy, press releases, testimony, social posts, and email to major donors. Anywhere AI supplies a number, a source, or a quotation, it needs checking before it goes out.
The verification workflow
The fix is a habit, not a tool. Every fact an AI gives you is unverified until someone on your team has confirmed it in a real source.
Assume every specific is unverified. Numbers, study names, quotations, funder details, eligibility rules, and dates are claims to check, not facts.
Trace each figure to its original source. Find the report, dataset, or publication that first published it. A blog post repeating a number is not the source.
Open the source and read the relevant part. Confirm the figure, the year, the geography, the population, and the definition all match how you are using it.
Check quotations word for word. Any quotation, whether from research or from a participant, must match the original exactly and be one you have permission to use.
Confirm funder details at the funder. Check deadlines, eligibility, award ranges, and required attachments against the funder's current published guidelines or the official funding notice, not the AI's summary of them.
Check your own numbers against your own records. Participant counts, outcomes, and budget figures should match your program data, database, and financial records.
If you cannot verify it, cut it. An unverifiable claim does not go in a proposal, report, or appeal. Replace it with something you can source, or leave it out.
Habits that reduce the risk
Verification is not optional, but these habits make fabrications less frequent and easier to spot:
Supply the source material. Paste in the funding guidelines, your program data, or the research report, and instruct the AI to use only what you provided. It still makes mistakes, but they are far easier to check.
Ask where each fact came from. Request that every figure be tied to a specific document and page or section, so a reviewer can go straight to it.
Make unverified claims visible. Tell the AI to insert a marker such as [VERIFY] or a placeholder like [STAT NEEDED] instead of supplying a number it cannot source. Nothing with a marker goes out.
Keep a source log. For each proposal or report, keep a short list of every figure with its source and who checked it. It also speeds up next year's renewal.
Use search-connected tools with care. Tools that search the web or your files tend to fabricate less than a bare chatbot, but they still misread sources. The workflow above still applies. See tool setup for configuring what your team uses.
The bottom line
AI is a real time-saver for drafting, restructuring, and tightening grant and donor copy. It is not a source of facts. Keep the drafting role and the fact-checking role separate, and the risk becomes manageable.
Before you submit: a checklist for proposals and reports
Every statistic has been traced to its original source, and the source has been opened and read.
Every cited study or report has been confirmed to exist and to say what the document claims.
Every quotation matches its source word for word, and participant quotes have documented consent.
Deadlines, eligibility rules, page limits, and required attachments match the funder's current guidelines.
Participant counts, outcomes, and budget figures match internal records.
No [VERIFY] or placeholder markers remain in the document.
Anything that could not be verified has been removed.
A named staff member, not the AI, has reviewed the final version and stands behind it.
Sources & further reading
OpenAI, "Why language models hallucinate" (September 2025) — openai.com
National Institutes of Health, NOT-OD-25-132, "Supporting Fairness and Originality in NIH Research Applications" (July 17, 2025) — grants.nih.gov