What your funder's AI says about you
23 diligence questions about a nonprofit I chair. AI had an answer for every one; about half the answers held up.
Dear Nonprofits,
Funders are asking AI about you. If that AI lacks a full picture of your work, it will fill in gaps from other sources, and the results can be unfairly negative.
I tested this on Vote Rev, an organization I co-founded and chair. I put three of its recent funder-facing documents into a leading AI model, turned on web search, and had the model answer 23 diligence questions on topics like evidence, cost-effectiveness, leadership, and “what would a sharp skeptic say.”
On about half of the 23 questions, AI's answer was worse than Vote Rev’s real story. Almost nothing AI said was false; like many bad AI answers, it was confident, plausible, and sourced, but often badly incomplete or off-base:
Asked about peers, AI called Vote Rev "not unique" and wrongly named other organizations as comparable.
Asked about cost-efficacy, AI searched the web and found inapplicable benchmark data, then dinged Vote Rev’s work as overly expensive.
Asked about leadership, AI couldn’t explain the founder’s departure, and scored that as key-person risk. The true story (an orderly transition, a national search, the board’s first-choice hire) isn’t laid out publicly and was out-of-scope for the pitch materials.
My biggest takeaway: Where the documents are silent, AI answers anyway, from whatever the web offers. (This is a common AI failure for subjects with thin coverage on the web.)
What to do about this
The solution is to create an easily AI-readable body of your materials to share with funders.
Step 1: Gather the documents
To pick which documents to include, imagine a large funder told you, “We hired a new person and their first job will be to spend a week getting up to speed on your organization. Send us everything.” You’d probably share things like:
Funder-facing materials (even if not related to that funder)
Annual reports and other periodic updates
External validation of your work
Conference presentations (slides and/or speeches)
Budgets and/or tax filings
Unlike a human, AI won’t be overwhelmed by too much information.
Step 2: Tweak the materials for AI
2a. Check for internal consistency
Ask AI, “Do these files have any inconsistencies, even minor ones that a skeptic might see as a red flag?” In my experience so far, Claude has been quite particular when it notices conflicting claims, even legitimate restatements or extrapolations. Explain or revise such points before your funder’s AI flags them.
2b. Tell your story clearly
Imagine that a few years ago, something bad happened at your organization. This “Bad Thing” should be included in your corpus documents, told honestly in the way you prefer. Wherever the corpus mentions the Bad Thing, insert a description of the resolution. Unlike human-facing materials, there’s near-zero cost to redundancy. Test this by asking your AI, “What do you make of [Bad Thing], and how has the organization responded?”
2c. Put key chart data in text alongside
AI reads text reliably, but can lose information from complex charts. If your materials have load-bearing exhibits, put the key takeaways in text nearby. (This should become unnecessary in the near future.)
Step 3: Run a final check
List your most common funder questions, then ask AI to add some that a skeptical funder might wonder but not say out loud. Give it those questions and your corpus, and tell it, “Answer each of these using the attached material and what you can find on the web.” Compare each answer to what you’d have said yourself.
The funders you haven’t met yet
The corpus you just built takes care of the funders you’re in touch with. Next you’ll want to do a similar exercise for your public presence.
Everyone wants AI to say nice things about them. Making this happen is a field so new it lacks a settled name; two common ones are Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO). With a field so young and models changing so fast, any given tactic may end up being short-lived. But the foundation will remain: make the true version retrievable. I’m not an expert but some basic tactics strike me as having little risk and good upside:
Put important facts where models are likely to search: your own site, your Candid/GuideStar profile, and your ProPublica Nonprofit Explorer page.
Keep the public footprint consistent with the private packet. For anything that you describe differently in private versus public materials, explain it in your private corpus.
State claims plainly, with data. “Our program costs $18 per registered voter, versus a field benchmark of ~$30” probably travels better than “cost-effective and rigorously evaluated.”
Conclusion
The theme running through all of this: where you are silent, AI answers anyway. The packet fixes that for funders you know; your public presence fixes it for the ones you don’t.
Today, I doubt many funders probe with AI as deeply as I do — so for now the ideas above are insurance, not an emergency. But soonish, I expect my behavior will be common. Eventually, funders’ and organizations’ AIs will exchange key information on their own. Until then, the solutions above are the best tools we have.
Closing question
If you run this test on your own organization, or if a funder’s AI has already gotten something wrong about your organization, I’d love to hear about it.