The third annual Donor Perceptions of AI study reached 1,020 U.S. donors in September 2026, bringing the three-year total to 3,057. Between 2024 and 2026, the share describing themselves as very familiar with AI rose from 31.6% to 49.7%. The share saying a nonprofit’s AI announcement would make them less likely to give moved from 31.4% to 32.3%. Discomfort with AI-personalized appeals stayed within two points of 40% in every wave.

Sampling is the obvious objection and it has an answer: the 2025 and 2026 samples are nearly identical in age, and familiarity still rose eight points between them while giving intentions did not move at all. Among donors very familiar with AI, 29.0% would still be less likely to give, and the expectation of disclosure is highest in that same group, at 85.0%. Sophistication raises the bar rather than lowering it.

Then the result that does not divide donors, which here is rare. Asked which AI applications concern them, 70.2% selected bots portrayed as humans representing a charity; the next one down drew 49.8%. Nathan cut the sample sixteen ways and a majority selects it in fifteen of them, from donors giving under $100 to donors giving over $5,000. Among those who said their giving depends on how AI is used, it reaches 85.4%.

This is actually much bigger than just AI. The assumption that just failed has a name. Science communication calls it the deficit model: resistance to a technology is a knowledge gap, so fill the gap. In Nature Climate Change in 2012, Dan Kahan, Ellen Peters, Maggie Wittlin, Paul Slovic and colleagues found the reverse. The people highest in science literacy and numeracy were not the ones most aligned with the evidence; they were the group among whom polarization was greatest.

Comprehension gave each side better tools for holding the view it had. Our study numbers carry that proof: the undecided share nearly halved, and those donors left in both directions, optimism gaining about five points and skepticism about four. So the question is not whether donors understand AI. It is whether your communications are asking them to take a side, because an explainer is an invitation to hold a position, and the better it is written the more cleanly it sorts.

That also explains why concealed identity is the line. The objection is not to a machine in the building; support chatbots drew 36.1% and predictive analytics 24.8%, both modest. It is to a machine wearing a name and a face. Reading a note from a charity is a social act, not an informational one: the reader runs a simulation of the person on the other end, the mirror-system machinery that makes any story land. In Neurogiving I spend a chapter on how giving tells a donor who she is, and that only works if what she measures herself against is real. A bot with a human first name does not fail because it performs badly. It fails because the simulation is running on nobody.

Stop making AI the subject of anything you send a donor. Not only the explainer and the webinar, but the newsletter line about exciting new tools and the innovation paragraph in the annual report. Each asks a supporter to form a position on a technology, which is the one move the evidence says hardens whichever way she already leaned. Your donors do not need a view about AI. They need to know who is accountable for what reaches them.

Two things to carry into the room this week.

The sentence is a disclosure you can say out loud without flinching. Three parts: what the software does, who checks it, how to reach a person. “Software helps us organize the reports our grantees send. Our program team checks the information and writes what you receive. If you want to talk about it, call me directly.” Read it again for what is absent. No claim about AI, nothing to agree or disagree with, no side to take. That is the design. A division of labor gives a donor something to judge; a position on the technology gives her something to join.

The question is for your own team: does anything we send a donor carry a human name, a face, a signature, or an opening line a reasonable person would read as a person? Not what it is called internally. What a donor would conclude. A named assistant is fine if it says what it is in its first sentence. A signature on an automated message is a different thing, and 70.2% of your file has a very strong view about that.

The Referendum. The organization decides to be open about AI and announces it: a page, a webinar, a confident line in the newsletter. What it has done is put a question to its donors that nobody asked them to vote on, and the better the explanation, the more cleanly the file divides. The spend is real, the intent is honest, and the effect is sorting. The corrective is to change the subject rather than the wording. “We are transforming our work with AI” is a referendum. “Software organizes the reports, our team writes the conclusions, here is who to call” is an accountability statement, and nobody has to pick a side to accept it.

Research Sources

In Lab Notes Pro

There is a number in the outside research that I have been arguing with for two weeks, because it appears to contradict the recommendation I just made to the whole sector. It is large, it is from a clean field experiment, and the obvious reading of it is that disclosure is expensive. I do not think the obvious reading is right, and working out why changed what I think the number is actually measuring. That argument, and what it costs me to make it honestly, is in Pro along with a board ready one-pager on these findings.