Nobody at your association applied for permission to start using AI. It arrived the way email arrived, one person at a time, because it saved an hour on a task nobody enjoyed. By the time it reaches a leadership agenda it is not a proposal, it is a description of how the work already gets done.
The question is not whether to allow it. The question is whether anyone has said what may go into it and what has to be true before the output goes out under your organization's name.
Where this sits. This is the fifth and last of five posts on AI and association websites. The survey figures are from the organizations that ran the surveys, and where the popular version of a story turned out to be weaker than its reputation, we have said so.
The Numbers for This Sector Specifically
NTEN and The Bridgespan Group surveyed 917 nonprofit staff and executives over the summer and published on September 10, 2026. It is the most directly relevant dataset available and it is three weeks old.
- Use is near universal. 74.42 percent use AI regularly. 45.37 percent use it daily or more often.
- Governance is not. 38.56 percent of organizations have written guidance on responsible use. 39.95 percent have established rules on what data can be entered into an AI tool. 30.42 percent have a formal vendor approval process.
- And people taught themselves. 78.03 percent of staff say they learned through hands-on experimentation. 36.82 percent report no training at all, and 57.21 percent of organizations have no budget for AI.
Put those side by side. Three quarters of your colleagues are using it. Two fifths of organizations have said anything about what data may go in. Nothing else in this post is as important as that gap.
And it is not just this sector. The broader picture matches. Cisco's 2024 privacy benchmark, a double-blind study of 2,600 security and privacy professionals across twelve countries, found 48 percent said non-public company information had been entered into generative AI tools, 45 percent employee information and 38 percent customer information. A University of Melbourne and KPMG study of 48,340 people across 47 countries found that almost half admit to using AI in ways that contravene organizational policies, and that over half do not reveal when they have used it.
The Cautionary Tale Is Thinner Than You Have Heard
The story everyone tells is Samsung, where engineers supposedly pasted proprietary source code into ChatGPT and the company banned it. It gets repeated as established fact and it is worth being precise about.
What is actually documented. The original report was an exclusive by Economist Korea on March 30, 2023, describing three incidents within about twenty days of the semiconductor division permitting ChatGPT use. It was based on internal reporting. A Samsung representative declined to comment. Every English-language account traces back to that one article, and Samsung has never confirmed the incidents. Bloomberg reported the subsequent ban on May 2, 2023 based on a memo it reviewed.
Why this matters for how you argue it. We are not saying it did not happen. We are saying that if your case for a policy rests on a story your staff can find out was never confirmed, you have a weaker argument than the one sitting in your own sector's survey data. Use the NTEN numbers.
The Rule That Fits on One Line
Most AI policies fail the same way: four pages, a long list of prohibited uses, written by somebody who will not do the work, read once. The best formulation we have seen is a single sentence, from the Administrative Conference of the United States policy of December 2025. Staff should not share information with a generative AI system that they are not permitted to share publicly.
Why that one works. It needs no data classification scheme, it survives being explained in a staff meeting, and it produces the right answer for almost every case someone will actually face. Member records, no. Board minutes not yet public, no. A draft press release you are about to publish anyway, yes. A public page you are rewriting, yes.
If you already classify data, hook into that. Harvard ties the same idea to an existing classification, prohibiting confidential data above a defined level in publicly available tools and permitting only systems assessed and approved by its security and privacy office. If your association already has a data classification policy, use it rather than inventing a parallel one. If you do not have one, use the sentence.
Two published policies worth copying from. For a fuller model written for this sector, NTEN publishes a generative AI use policy template. The Manhattan Borough President's office publishes a short operational one that names prohibited categories, requires independent verification of any AI-produced fact before public use, and requires citation of the tool in public-facing materials. Both are real documents you can read rather than frameworks you have to buy.
What Google Actually Requires, Which Is Less Than You Think
A recurring worry is that publishing AI-assisted content will be penalized, or that it must be labeled. Google's own documentation does not say that.
On AI content generally. Its spam policy targets scaled content abuse, defined as many pages generated primarily to manipulate rankings rather than help users, and states this applies no matter how the content is created. The first example given is using generative AI to produce many pages without adding value. The test is purpose and value, not production method.
On disclosure specifically. Google's helpful content documentation asks whether the use of automation is self-evident to visitors through disclosures, and concludes that AI or automation disclosures are useful for content where someone might think how was this created, and to consider adding these when it would be reasonably expected. That is a recommendation keyed to reader expectation, not a requirement.
And no state requires labeling either. And no state law requires an association to label AI-assisted marketing copy, web content or a newsletter. The state laws that exist govern interactive systems, mainly chatbots, and regulated services. Disclosure of AI-assisted content is a judgment call your organization owns, not a compliance box, and we think most associations should make it for reasons of trust rather than rules.
The Argument That Works Better Than Fear
There is a reason to require human work on AI output that has nothing to do with risk, and executives respond to it better.
You do not own what the machine wrote by itself. The United States Copyright Office has been consistent: copyright protects only material that is the product of human creativity, and purely AI-generated material is not protectable. Its January 2025 report states that prompts alone do not provide sufficient control, and that human authors are entitled to copyright in their own expression perceptible in AI-generated outputs, as well as the creative selection, coordination or arrangement of material and creative modifications of the outputs. The DC Circuit held in March 2025 that the Copyright Act requires a human author.
For an association that is not abstract. Your association publishes standards, model policies, curricula, survey reports and position papers. Some of that is licensed, some of it is sold, and all of it is the reason people join. Material generated without meaningful human authorship is material you may not be able to protect. That reframes the review requirement from a rule about risk into a rule about owning the asset, and in our experience that is the version that gets adopted.
Four Things to Put in Writing
Short enough that people read it, specific enough to answer the cases that come up:
- What must never go in. The one-line public-shareability test, plus a short explicit list for the things people get wrong: member records, unpublished board material, anything under an NDA, personnel matters.
- Which tools are approved, and who approves a new one. One named person approves new tools, with the vendor, the intended use and the terms recorded. A third of organizations have this; it is the cheapest item on the list.
- What has to happen before publication. Any fact, figure or citation produced by AI is verified against a non-AI source before publication, by a person with the expertise to judge it. Not because of hallucination statistics, but because it is your name on it.
- Whether and where you disclose. Not a compliance requirement, a trust decision. Decide it once, write it down, apply it consistently, and tell your members what the rule is.
Write It This Quarter, Not This Year
This is the one area in this series where the work does not need us. The published policies above are free, the one-line rule is genuinely sufficient for most organizations, and the honest first version is two pages written by whoever will have to enforce it. Doing that badly this quarter beats doing it properly next year, because your staff are making these calls on their own in the meantime.
Where we do help is the part that touches the website: what your content standards should say now, how AI-assisted material moves through review before it publishes, and whether the disclosure you decide on belongs on the page, in the byline or nowhere. If you want a second read on a draft policy, send it and we will tell you what it does not cover, which is usually contractors and agencies. That has been the gap in every draft we have read this year, including an early version of our own.