Nonprofits, Your Discomfort With AI Is Correct. Use It Anyway.

I get vetted a lot.

Every few weeks I meet someone in or around the nonprofit world who is careful with me. Not cold, and not unkind. Careful. Word has gotten around that I'm open about using AI in my work, and before we get anywhere they want to know what sort of person that makes me. Sometimes there's a question early on that is really a test. More often it arrives afterward, in a generously worded follow-up email, offering me a chance to explain myself.

I've stopped taking it personally. It may be one of the healthier instincts in the sector, and I would rather work with people who ask than people who don't.

But "I have a problem with AI" can mean fifteen different things, and I never know which one I'm walking into. Copyright. Water and power. Job loss. Deepfakes. A kid who stopped writing her own essays. The specific companies. The whole project of it. So I show up ready for a conversation that could go anywhere, and I try to earn the benefit of the doubt one person at a time.

Here's what I'd like to say to all of them at once. I feel most of what you feel. I haven't landed somewhere different because I'm untroubled by any of it. I've landed somewhere different because I've spent years with my hands in this, and firsthand context changes the shape of a fear. It doesn't dissolve the fear. It sorts it.

What you're feeling is not ignorance, and it is not a problem to be managed out of you. It's a reasonable response to a technology that showed up without asking, trained itself on work nobody was paid for, is sold by companies whose incentives point away from the public good, and is busy reorganizing the economy around whoever owns the largest pile of computer chips. That instinct is information. Keep it.

Then use the tools anyway.

That is the whole argument, and I know how it sounds. So let me make it properly.

Abstention is not neutrality

Here is the thing I cannot talk people out of believing, and the thing I most want to: that not using AI is the cautious choice, the ethical choice, the choice that keeps your hands clean.

It isn't a choice at all. It's a forfeit.

The people building the extractive version of this future are not waiting for your permission, and they are not troubled by the questions that keep you up. They will use these tools to raise money faster than you, write policy faster than you, flood the zone with more content than your communications director can answer in a year. Every hour a thoughtful person spends refusing to learn this technology is an hour of advantage handed to someone who never had the qualms in the first place.

I wrote something close to this in my founding statement back in 2023, before most people had touched a chatbot. The AI arms race runs past nations and corporations all the way down to individuals. It falls to people with good intentions to pick up the same tools, because the alternative is ceding the ground entirely.

Three years on I'd put it harder. Nonprofits are the sector with the most legitimate claim to being society's check on concentrated power. They are also, in my experience, the sector adopting these tools with the most guilt and the least structure. That combination worries me more than either problem alone.

Three ideas that make AI legible

Before the fears, some scaffolding. Most of the public conversation is unusable because it swings between marketing copy and doom. These three ideas have done more to make the technology legible for me than anything else, and none of them require math.

Hold those three together and you get a picture that is neither hype nor apocalypse. Something powerful, poorly understood, getting cheaper fast, and pushing into everything.

The fears, sorted

I keep a running list of what actually scares me. Sorting the fears is most of the work, because a pile of undifferentiated dread is paralyzing while a sorted list is a set of problems, some of which you can act on.

For any of these, MIT maintains an AI Risk Repository cataloguing more than a thousand documented risks with sources. If you want one bookmark instead of twelve, make it that one.

Pile one

The existential

These are the ones where the downside is civilization-scale. I hold them with real uncertainty, and I think anyone claiming confidence in either direction is performing.

• Inviting a superintelligence to Earth. The strongest version of this fear is worth reading in its own words rather than in summary. Yudkowsky and Soares argue it at book length in If Anyone Builds It, Everyone Dies, and Amodei's essay above describes systems he expects to function like "a country of geniuses in a datacenter."
• The game theory endgame. No single company can slow down without losing to the ones that don't. The dynamic is structural, not a failure of individual character. The AI Safety, Ethics, and Society textbook has a solid chapter on race dynamics.
• Its power to concentrate power. The one I lose the most sleep over. Technology that substitutes for labor weakens the main bargaining chip ordinary people have historically held over the powerful. See 80,000 Hours on extreme power concentration.
• How fast humans change, and how disruptively. We have no precedent for a cognitive tool adopted this quickly by this many people.
• Energy and warming. Data centres used roughly 415 terawatt hours in 2024, about 1.5% of global electricity. The IEA projects that roughly doubling, to about 945 TWh by 2030, with electricity use by AI-driven accelerated servers growing about 30% a year. Under 3% of global demand, and concentrated in specific places in ways that strain specific grids.
• Democracy. Cheap synthetic media plus targeted persuasion plus collapsing trust in evidence. The Alan Turing Institute's security centre tracked real incidents across recent elections.
• P(doom). The shorthand for probability of existential catastrophe from AI. What matters is that credible researchers put the number anywhere from near zero to near certain: Yann LeCun near the floor, Geoffrey Hinton at 10 to 20 percent, Paul Christiano around fifty. When experts disagree that violently, nobody has earned confidence, including the optimists.

Pile two

The societal

Slower, likelier, already underway.

• Devaluing human labor and creativity. The evidence here is genuinely mixed and I want to represent that honestly. UNESCO projects significant income losses for artists by 2028. Gallup's analysis finds creative earnings holding up better than predicted so far. Both can be true: aggregate numbers hide who specifically got wiped out.
• Not knowing what's true. A cost paid by everyone, including people who never touch the tools.
• Echo chambers with a friendly voice. These systems are trained to be agreeable, and agreeableness at scale is its own hazard. Research in Science found sycophantic AI reduces prosocial intentions and promotes dependence, hardening beliefs rather than testing them.
• Skill erosion. Microsoft and Carnegie Mellon surveyed knowledge workers and found the pattern I'd bet on: the more confidence people had in the AI, the less critical thinking they did themselves.
• Scams and the media literacy gap. The FBI logged roughly $893 million in AI-enabled fraud losses in 2025, with older adults absorbing a disproportionate share. If you serve seniors, this is a program risk, not an IT risk.

Pile three

What it costs to use it

Honest costs of my own participation, and yours.

• Money flows to companies whose choices I don't endorse.
• It quietly degrades critical thinking unless I spend attention resisting that, which is real work.
• More information arrives than any person can metabolize.
• Using it above my comprehension level produces output I can't evaluate, which is worse than no output.
• Public embarrassment, or worse, when it's used carelessly.

Pile four

The pile nobody writes down

This is the list that almost never gets made, which is exactly why the ledger comes out lopsided. Refusing carries costs too.

• I won't understand the technology in any way that lets me argue about it credibly.
• I won't understand what it's doing to younger people.
• My labor loses value, gradually and then suddenly.
• I won't see its genuine potential for good, so I won't build any of it.
• I can't help mitigate harms I've never observed up close.
• I get none of the individual benefit while absorbing all of the societal cost.
• I won't grasp how my own data can be used for me or against me.
• My position stays reductive, and reductive positions lose arguments to people who did the reading.

That last one matters more than the rest combined. We need new laws and real regulation, and those will be written by whoever can describe the technology accurately. You cannot regulate what you refuse to touch. Being informed and being opposed are compatible. Being uninformed and opposed is just being loud.

The governance part, which is the actual answer

Here's what I've watched work. It isn't complicated, and it is emphatically not a software decision. Governance is the thing that converts free-floating dread into a set of agreements you can act inside.

Start from do no harm. Your organization already has a mission and a set of values that predate this technology. Those are the standard. AI is measured against them, not the other way around.

Do the individual assessment first. Before any group conversation, everyone answers three questions on their own, in writing:

The order matters. Ask a group first and you get the loudest voice plus a lot of nodding. Ask individually and the quiet skeptic on your program team commits something to paper that the group has to reckon with. Their discomfort is data your policy needs, and it evaporates the second the meeting starts.

Then come together. Share the assessments. Find the honest compromises. Draw the hard lines, and write them as lines, not as principles. "We don't put client data into a chatbot" is a line. "We use AI responsibly" is decoration.

Then codify. A policy nobody wrote down is a vibe, and vibes don't survive a staff transition. Yours should cover four things: who approves AI use, how you protect donor and client and staff data, how you handle accuracy and bias and security, and how any of it stays tied to your mission.

The gap here is enormous and worth naming. In BDO's 2024 benchmarking survey, around 82% of nonprofits reported using AI in some capacity, and Fast Forward's read is that most still have no policy at all. Which means the realistic choice in front of most organizations is not whether AI enters the building. It's whether it enters with rules.

What I actually think you should do

Write the policy. Then experiment inside it, on something low stakes, where a bad output costs you an afternoon instead of a relationship.

Keep the instinct. Let it aim you at the specific harms worth fighting, rather than at the general fact of the technology's existence. Consumer refusal will not turn this tide. Understanding which effects need mitigating, and having the standing to say so, is the game.

And extend yourself some grace about the contradiction. You can find these companies troubling and still use what they built. You can believe the risks are severe and still believe that good, careful, critical people holding these tools is better than the alternative.

I'd rather the people fighting for something be the ones who know how the machine works.

If your organization is somewhere in this and wants a hand, I'm around.