— KEY TAKEAWAYS
- AI is safe for nonprofits when you use in-product tools with zero-retention agreements and avoid pasting donor data into personal or free accounts in consumer tools like ChatGPT or Claude.
- Your donor data does not “train” AI models unless you feed it to a public tool account that doesn’t have the proper protections and guarantees.
- Effective, safe use comes down to three tactics: keep donor data out of free tools, always fact-check AI outputs, and treat AI like a supervised intern rather than an all-knowing oracle.
- A common mistake is assuming AI “knows” the answer to a specific question about your own data. Without access to your data, it will hallucinate a plausible-sounding but false answer.
AI is everywhere these days. Whether you’re choosing a show to watch on Netflix or spinning up a rough draft of an email, the technology we use every day is getting smarter and changing the way we work.
But lots of nonprofit professionals have been asking questions about how they can use those tools safely without compromising their supporters’ data or jeopardizing the personal relationships with them they’ve worked so hard to build.
With those questions in mind, we hosted a live webinar where nonprofit professionals asked an actual AI product manager the questions they’d been too nervous, too busy, or too confused to ask. Samuel Chen, who leads AI Intelligence at Neon One, sat down with us to explain how these tools actually work, when you can (and shouldn’t) use them to understand donor data, and some steps you can take to keep your community and your staff safe.
Here are some of the high points of the interview.
What Are the Different Kinds of AI, Anyway?
The word “AI” gets used for a lot of different things these days, and that’s part of why the topic can be so confusing. There are really two types of AI.
The first is machine learning. Think about a major gifts officer with 20 years of experience who can look at a donor record and instinctively know when a person is ready to be asked for a larger gift. They can’t always explain why they know a donor is receptive—they just know.
That intuition, built from thousands of donor interactions over the course of their career, is the human version of what a machine learning model does. Machine learning can look at years of historical data, learn the patterns, and predict things like who’s at risk of lapsing or what a good ask amount might be.
The second type is what most people actually mean when they say AI today; it frequently refers to chatbots like ChatGPT, Claude, and Gemini. Those tools run on large language models (LLMs), and they’re also prediction machines.
LLMs don’t actually “think” or reason—they just predict the next word in a sequence. Those predictions are based on everything they’ve read (and they have been trained on incredibly large amounts of text) and everything you’ve included in your prompt.
It’s like the autocomplete feature on your phone. Imagine that autocomplete has read nearly everything humans have ever written. That’s an LLM!
Wait, Is My Donor Data Training These Tools?
This might be the most important question of the whole session, so here’s the direct answer: no. Unless you’re giving all of your donors’ data to LLMs, it’s not training anything.
There are two ways to give an AI tool knowledge. One is to train it by feeding it data that becomes permanently baked into the model’s memory. The other is what Neon One does, and it’s a little harder to explain; a good analogy is that it’s like an open-book exam instead of a closed-book exam.
Did you ever take an open-book test when you were in school? You read a question, look up the answer, and write it down. There’s no memorization, and you close the book and move on at the end of the test.
Tools like the ones we’re building in Neon One (shhh, spoiler alert!) are similar. When you ask a question, the system takes an open-book test. It pulls the relevant records from your secure database, shows them to the model at that moment, and the model gives you an answer to your question.
Then, the exam is over. The book closes. Nothing gets memorized, and the next question starts fresh. Your data lives in your database the whole time, and it never becomes part of the model itself.
On top of that architecture, there’s a legal layer. Neon One only works with AI providers through enterprise agreements that include a zero retention guarantee. That means the provider is contractually barred from logging or training its models on any of your donors’ data.
There is an instance where your donor data might be training a larger model, and it’s something to keep in mind if you’re using free or personal versions of tools like Claude or ChatGPT. If you paste donor information into a free consumer AI tool, that information may be stored and used to train the model going forward.
If you’re using an enterprise or business LLM or other tool, you can protect yourself and your donors by changing the settings on your account. Go into your settings and turn off the option that lets the system use your data for training. If a tool you’re using is already built into a platform like Neon One, you shouldn’t have to take this step.
How Neon One Approaches AI (and What It Means for Nonprofits)
Why Do AI Tools Make Things Up? How Do I Catch Those Mistakes?
Sometimes, AI tools generate information that sounds true but isn’t. Those mistakes are called hallucinations.
Remember, these models work by predicting the most plausible next word in a sequence. But “plausible” doesn’t always mean “true.” Models don’t actually think or reason. They were built to produce text that sounds right, and they’re correct most of the time.
But when a model doesn’t know something, it usually won’t tell you it doesn’t know. Instead, it will simply guess what word might come next. That’s why models can invent things like a donor’s name, a fundraising total, or a statistic that sounds completely believable.
An example of this you might remember is when Google partnered with Reddit in 2024 to train their AI search model on the site’s user’s posts. This resulted in the AI misunderstanding a joke post and guessing that “add glue” was a plausible step to include in “how to make a pizza.”
LLMs have advanced a lot since then, but still happen, and they are more likely to happen when a tool gets a very specific question without the ability to find real answers.
If you ask an AI tool like ChatGPT what makes a good year-end appeal, you’ll get a pretty solid answer. It’s read thousands of them! But, if you ask it how much your fall gala raised last year and it doesn’t have access to your actual data, the model has no way of knowing the correct answer. It will just invent something that sounds plausible.
There are two ways to prevent hallucinations.
The first is to use tools built into a product like Neon One. Because those tools are built directly into the source of all your data, the model never has to answer questions about your data from memory.
Instead, it will run an actual query against your nonprofit database and report what it finds. It never has to make up a plausible–sounding answer to your question because all the information to give you an accurate one is immediately available.
The second is to be very explicit when building prompts. Be specific about what you want, and give the tools you’re using enough data to make an informed decision (just remember not to give it donor data!). Once you have an answer, do some fact-checking to make sure everything’s accurate.
Nonprofit Data Management 101: What You Need to Know
Treat AI Like a New Intern, Not an Oracle
A useful way to approach all this is to treat any AI-powered tool like an intern.
If a talented new intern made a mistake in their first week, you wouldn’t fire them and swear off interns forever. You’d spend a little time learning what they’re good at, where they need help, and how to give them clear direction.
That’s a really valuable way to handle AI tools, too.
You can delegate some responsibilities entirely to your intern, give them some jobs that you’ll keep a close eye on, and keep some tasks entirely to yourself. The same applies with AI tools. You’d never let an intern send something to your biggest donor without reviewing it first, for example, but you might let them write the first draft. The same concept applies to an AI tool.
Direction matters just as much to AI tools as it does to interns. If you tell a new intern to “write something for your newsletter,” you’ll probably get something really generic. But if you give them a little guidance—if you tell them to write 200 words for a monthly donor newsletter about a specific program update, ask them to use a warm tone, and have them use the terms your organization actually uses for its donors—you’ll get something usable.
The quality of what your intern gives you is largely determined by how much direction you give them up front. The same thing is true with an AI tool!
And, if you ever get frustrated (which you probably will), remind yourself that this is the worst the technology will ever be. Models improve every few months!
5 Times When Nonprofits Should NOT Use AI
Where’s the Real Opportunity for Nonprofits?
Nonprofits can use AI tools to do tedious work so they can focus on human-to-human work. That’s the biggest opportunity here.
Many organizations are sitting on years of rich donor data without a dedicated analyst to mine it. If that’s the case at your nonprofit, the amount of time you spend running reports and understanding that data means you have less time to spend on thank-you calls, personal notes, and other little interactions that keep your donor relationships strong.
The goal for your AI tools should largely be to reduce tedious work, so you have time to spend on those meaningful activities.
AI Is Safe for Nonprofits, But Results Depend On You
AI is safe for nonprofits, but there are a few things you’ll want to keep in mind when you use it.
The first thing to remember is that your donor data should never go into personal accounts on public tools like Claude or ChatGPT. If you’re using AI to analyze data, make sure you’re using in-product tools (like the ones we’re building in Neon One) or are using paid business accounts that let you opt out of data sharing.
The second thing to keep in mind is that AI tools don’t reason. They predict. That means that any “output” you get from one of those tools must be fact-checked instead of being taken at face value. Write good prompts, make sure you’re sharing the relevant information, and double-check that the results you get are accurate.
The third and final fact to remember is that no AI tool will ever replace you or your team. Can they free up your time by simplifying your routines? Yes. Can they do tedious work? Yes. Can they help you speed up tasks? Yes. But you should think of those tools as interns that help you with your work. They’re going to require supervision!
Want to Learn More?
If you’re looking to understand more about how AI tools work, how they’re built, and how you can use them, check out this webinar. It’s full of useful information that will help you get the most out of the tools you use!
