AI for grant writing without inventing facts

A workflow for using an AI assistant on proposals while keeping invented statistics, fake citations and false organizational claims away from your funder.

WhatsAppEmail

What You Will End Up With

A grant draft in which every statistic, citation and claim about your organization traces to a source document you actually hold, plus a claim-to-source record a reviewer can check.

The reason for the effort is that language models hallucinate: they produce false or unverifiable material that reads as coherent and credible. In one experiment, a widely used chatbot produced 176 citations for six literature reviews, and 35 of them (19.9%) were fabricated. Of the real ones, 45.4% still had errors, most often wrong or invalid DOIs. A larger study found that 3% to 13% of citation URLs from commercial models and research agents were hallucinated, and 5% to 18% did not resolve at all.

None of this means the tools are useless. Generative AI goes beyond automating tasks: it creates new content, and a national charity network lists tailored grant applications among the tasks where it can help small nonprofits. The same network names data protection, cyber security and misinformation as risks. The workflow below keeps the benefit and puts a check on each of those risks.

Before You Start

Settle these before you open a tool:

  • The funder's current AI guidance. Some funders allow AI use if you disclose it, and others restrict it.
  • Your own verified material. Program records, data you can cite, and the claims you can back. Start from your own words and ideas.
  • An organizational AI policy. It should cover governance, data privacy, risk management and ethics, including how you handle accuracy, bias and security.
  • The approved account. Some organizations require team or enterprise accounts, where settings can limit how data is retained and used.

Steps

  1. Check the funder's rules. Do this first, for every funder, since guidance differs and changes. If disclosure is required, plan where it goes in the application.

  2. Write your own material first. Collect the data and program details, then draft in your own words. One guide says reviewers can often tell when grant language is generic or detached from the applicant's voice.

  3. Give the assistant one short section at a time. Assistants give better feedback on shorter text. Say which application and which section the text serves. For example: "I am writing a proposal for [application]. The text below is for the [section] section. It is my own draft."

  4. Ask for tone, clarity or grammar, not new content. For example: "Check this text for clarity, grammar and tone. Do not add facts, numbers or references, and do not change what any claim says." Then read the result against your original by hand. Look for anything new, especially a plausible statistic or reference slipped in to sound authoritative.

  5. Verify every fact, reference and claim. Check all of them, not a sample. Assistants can make up sources, citations and data.

    • Search several scholarly databases and publisher sites for the cited title, as the researchers in the citation experiment did. For a report or statistic that scholarly databases are unlikely to hold, check the publisher's own site too. If you still cannot find it, treat it as fabricated and leave it out.
    • Open the link or DOI, but do not stop there. A fabricated reference can carry a working DOI that leads to an unrelated article, so read the source and confirm it supports the exact claim.
    • Open every link. One that returns an error or fails to connect does not resolve, and if it also has no archived record of ever existing, it was probably hallucinated. Either way, a link that does open still has to be read.
    • Fact-check any idea the assistant brainstormed before it goes into the draft.
    • Test unfamiliar tools on content you already know well, so you learn where they fail.

  6. Have a person review, and record it. Ask a colleague who was not involved in the prompting to read the draft. People catch what AI cannot, such as cultural tone and strategic fit. Keep a claim-to-source record: one row per statistic, citation or organizational claim, with the source document it came from and who checked it. A responsible-AI framework for fundraising advises documenting review processes and audit trails.

  7. Write down how you used the tool. Be transparent about where AI touched the draft, in line with the funder's disclosure rules. A responsible-AI framework for fundraising says any AI system that acts on critical decisions without human involvement needs ongoing human review and authorization. Apply the same idea here: a person reviews every passage the assistant touched before it goes to a funder. The drafting itself stays with you: use the assistant to polish, and rely on your own material for content.

Using AI Safely

Public chatbots may store what you type to update future models, which can expose your original ideas to other users.

  • Do not paste unpublished data, budget details or sensitive personal or organizational data into public tools. Use a secure tool your institution provides.
  • Protect the personal and sensitive data of donors, beneficiaries and stakeholders.
  • Use team or enterprise accounts where your organization requires them.
  • Check each returned item against its source, however good the tool. Bias in training data can bias output, so staff need training as well as rules.

A short AI policy helps here. One benchmarking survey cited by a nonprofit accelerator found almost 82% of nonprofits already use AI, yet most have no policy. Put yours in writing before the next deadline, not after a problem. AI use is not risk-free, and it requires intention, training, oversight and transparency. Name accuracy and bias in it, next to the risks above.

Common Mistakes

Trusting a link because it opens. It may resolve to an unrelated article. Read the source.

Believing more citations means better ones. More citations per answer does not make each one more reliable. Research-agent tools that produce long, heavily cited reports had the highest hallucination rates in the study.

Relaxing on niche topics. Fabrication ran at 28% to 29% on two niche topics against 6% on a well-known one. The more specialized your subject, the more you check.

Letting the assistant write. Draft yourself first and use the tool only to polish.

Testing only on unfamiliar material. You cannot judge output on a subject you do not know. Run a trial first, then use it on a live proposal.

Checking a sample. Verification means checking every citation and claim systematically, not a few. The authors of the citation experiment call for rigorous human verification of every model-generated reference, and a partial check leaves the fabricated ones in.

Skipping the record. Without a claim-to-source record, you cannot show a reviewer or a funder how a statement was checked. Keep it as you go, not at the end.

Pasting sensitive data into a public tool. Move the work to a secure account, or leave the data out.