Somewhere in the second hour of a meeting about an AI policy, after we’d gone around the table and everyone had named the thing they were worried about, somebody said the reasonable thing.
“What if we just don’t?”
Nobody jumped on it. There was a pause I’d describe as relieved. And it is a clean position, that one. No vendor review. No training line item. No uncomfortable conversation eighteen months from now about the thing the tool got wrong and who signed off on it. It sounds like caution, and in our sector caution reads as integrity.
I didn’t argue. I’ve been in enough of these meetings to recognize the moment when arguing costs you the room.
But I’ve been chewing on it , and on my lip, ever since, which is why I had to write this.
First, the part where I agree
I’ve spent a lot of this newsletter making the case against.
The energy and water these systems pull, and the fact that the people who host the data centers are rarely the people who benefit from them. Training data nobody was asked about. Bias coming back out of the machine wearing a lab coat. The whole shape of the thing, which looks an awful lot like patterns our field should recognize by now, and which I’ve been rude about in print.
None of that gets filed away because I’m about to argue the other direction. I’m not asking anybody to soften the critique.
I’m asking a narrower question, which is what the critique actually buys an organization once it gets converted into a policy that says no.
Because I think two different things have gotten fused together in our heads. Refusing to use a technology. And refusing to be affected by one.
Saying no doesn’t turn it off
A no-AI policy does not give you an organization without AI. I have never once watched it do that.
What it gives you is an organization where the AI use goes quiet.
Your staff are already using it. Not the reckless ones. The conscientious ones. The program manager with four deadlines and no research assistant, at 9:40 on a Tuesday night, on her own phone, on her own account, pasting in whatever she needs help with. The policy did not stop her. It made sure she can’t ask anyone for help doing it more carefully, and that she’ll never mention the time she pasted in something she shouldn’t have.
Then there’s everything you bought. It showed up in the grants portal, the case management system, the scheduling tool, the applicant screener. It arrived in a release note. Four bullets, and the third one was the AI. Nobody asked the board. And I’d guess a fair number of the organizations with the strongest positions on this have three or four vendors running client records through an AI model right now, often somebody else’s model on somebody else’s servers, because a checkbox got flipped on by default and nobody was told.
The question to ask your vendors, if you ask them nothing else: which of your features call an AI model, whose model is it, and where does our data go when they do? Most of them can answer. Almost nobody asks.
And your funders are using it. So is the person reading your proposal, and the agency scoring your application, and the consultant benchmarking your outcomes against your peers. Decisions about your organization are being made with these tools whether or not anyone on your staff ever opens one.
So the “no” isn’t buying you abstention. It’s buying a fairly expensive kind of not-knowing about what’s already happening in the building.
I’ve argued for refusing things before
Here’s the part that’s hard to say to people I mostly agree with. Your organization declining to use AI does not shrink a data center. These harms are structural. They move when procurement standards move, when disclosure gets mandated, when enough institutional buyers act together. They do not move because one nonprofit in Minneapolis decided not to use a chatbot. I wish they did. That would be a much easier world to organize in.
Which puts me in an awkward position, because if you’ve read this newsletter for any length of time you know I’m the last person who should be lecturing anybody about refusal.
I’ve walked away from a contract and written about why. I’ve argued that the most important thing an evaluator can do is sometimes know when to stay away, that when your presence is what creates the risk, absence is the intervention. I’ve written that rest is resistance, that exhaustion-as-currency is a colonial inheritance we can decline. I’ve told people to retire indicators, to stop measuring things, to say no to funders.
So I owe an honest account of why this refusal is different from those ones. I’ve been turning it over for a couple of weeks and I think it comes down to a single question.
Does your refusal withhold something the system actually needs?
Walking away from that contract worked because they wanted me. My participation was an input, and taking it back cost them something and cost me something, and that cost is precisely what made it a moral act instead of a preference.
Staying away from a community’s organizing worked because my presence was the risk. The documentation was the exposure. Absence reduced harm directly, immediately, in a way I could name.
Refusing the grind works because the grind runs on our consent. Withdrawing the consent is the whole mechanism.
Now run it on this one. What does a mid-sized nonprofit’s refusal withhold? Not compute. Nobody needed your queries. Not legitimacy, because this industry is not sitting around waiting for the social sector’s blessing. Not your data, which they either already have or never wanted.
Refusal is a precision instrument for situations where you are a necessary input. This isn’t one of those. And the single thing the refusal does successfully withhold is your judgment, from the rooms where the norms in your own field are getting set. Somebody is going to write the rules for how this gets used in evaluation, in service delivery, in grantmaking, and every ethical opt-out is one fewer ethical objection in that document.
Which is a strange outcome for a principled act. We kept back the one thing anybody could have used.
The gap won’t sort us by how careful we were
The other reason I can’t get comfortable: opting out doesn’t stop it for anybody else. It just picks which side of a widening gap you stand on, and I don’t think that gap is going to sort people by how thoughtful they were.
It’ll sort by capacity. Who can answer a question in an afternoon that used to eat three weeks. Who can read all of their qualitative data instead of a sample and a prayer. Who can turn around the thing the funder needs inside the window the funder gave.
And the organizations most likely to opt out are the small, community-rooted, chronically under-resourced ones. The ones whose caution is completely rational, because they have the least room for a public mistake. The ones with the most to gain from the work getting cheaper.
I wrote a while back about how much of our sector’s risk tolerance gets set by whoever happens to be in the room, and about the cost that never lands on anyone’s ledger — the decision made without evidence that existed the entire time. Same ledger here. On one side, harms we can picture: the bad output, the exposed record, the automated judgment nobody caught. Real. Worth being afraid of. On the other side, the analysis that didn’t happen, the program that didn’t improve, the community that went on being underserved because the organization serving it fell three years behind everybody else.
Nobody writes that one up. There’s no incident report for a slow decline.
You can’t govern a tool you’ve never watched fail
But the argument I actually care about is the safety one, and it’s why I keep showing up to these meetings in addition to writing think pieces about them. Which is, yes, what this is.
Last year I gave a model a set of interview transcripts I’d collected myself, from a project I knew cold, and asked for a summary. What came back was clean, well organized, and had quietly dropped the one participant who disagreed with everyone else.
That participant was the finding.
I caught it because I’d done the interviews. Somebody who hadn’t would have shipped it. And that is not a fact you can learn from a webinar. You learn it by using the thing on work you know well enough to check, and getting burned in a way you didn’t predict, and developing the small twitch that says look at that part again before you can explain why.
Governance from below needs people below with their hands on it. Otherwise our AI policies end up like most policies: written by people describing a system they’ve never operated, in language that sounds responsible and prevents nothing. I’ve written that language. I can recognize it now mainly because I’ve watched what it fails to stop.
Fluency isn’t a prize you get after you’ve settled the ethics. It’s the equipment you need to have the argument at all.
What I’d actually put in the policy
Not enthusiasm. A document that says yes with conditions and means both halves of that. Name the data that never goes in. Name the decisions a model never makes about a human being. Eligibility, termination, who gets the slot. Sort your uses into the ones that are fine unsupervised, the ones that need a second set of eyes, and the ones that need a conversation before anybody starts. Require people to say when they used it, so the use becomes visible instead of covert. Put somebody’s actual name next to each system. Put a review date on the whole thing, because anything any of us writes about this in August will be partly wrong by spring.
And keep the skeptics inside. Give the person with the sharpest objections a seat and a login, not an exemption. An objection from somebody who uses the tool every week is worth about ten from somebody who won’t touch it.
That’s a harder document to write than simply “no.” It’s also the only one that describes something true about your organization.
But know why the line is where it is
Those red lines are where I’d start. I want to be careful about them anyway, because “no model decides about a person” can be a principle or it can be a flinch, and from the outside the two look identical.
Here’s the uncomfortable part. The evidence on whether these systems are more or less biased than we are does not line up neatly behind either camp.
A 2024 audit out of the University of Washington ran more than three million résumé comparisons through three models and found they favored white-associated names 85 percent of the time, against 9 percent for Black-associated names. Black male names never once ranked above white male names. Damning, and roughly what most of us would predict.
Then a separate audit put 801 real applications for K–12 teaching jobs through eleven models and found modest disparities running the other direction, slightly favoring women and non-white candidates, at a magnitude the authors placed inside the normal range for human recruiters.
Both are careful studies. They disagree because they tested different models on different tasks with different prompts and measured different things. Which is itself the finding. Is AI biased is not a question that has an answer. Is this model, on this task, compared to what we are doing right now is a question that has an answer, and it’s one you can go get.
That last part is what I think our field should sit with. You can run three million comparisons through a model. You cannot run three million comparisons through your hiring committee. Human bias in these decisions is real, thoroughly documented, and not able to be audited at the moment it actually happens. Model bias is real and auditable. For people whose whole job is working out whether something does what it claims, that asymmetry ought to be interesting rather than beside the point.
The counterfactual is never a fair process. It’s the process you’re running today.
None of which means hand over the eligibility decisions. The failure mode on this side is real and it’s about consistency. A person’s bias is erratic and lands unevenly. A model’s is uniform and lands on everybody. One bad rule ten thousand times is a different animal than a hundred bad Tuesdays.
So keep the red lines. Just hold them as things you’d be willing to test rather than things you already know. We ask programs for evidence that they work all the time. We ought to be able to produce some for our own caution.
(This is roughly the structure I’ve been building a workshop around — the decide, implement, verify move, run against the ways these systems actually break rather than the ways we imagine they might. More on that when I have a date to give you.)
I’m not a neutral party
Obvious thing worth saying out loud: I’m fluent, and I make part of my living in a world that values fluency. When I tell you the only way through is practice, I’m arguing for the value of a skill I happen to have. That should cost my argument something with you. It costs it something with me.
I also know exactly how this reasoning goes bad. “We have to learn by doing” is about one bad quarter from becoming a permission slip for whatever somebody already wanted to do. The learning frame is only honest if it comes with a condition under which you’d stop.
So here’s mine, plainly enough that you can hold me to it. If the community conversation starts getting replaced by the summary of the community conversation. If the analyst’s judgment gets replaced by the analyst’s prompt. If we find ourselves believing outputs we can no longer independently check. That’s not learning anymore. That’s automating, and calling it learning because it’s cheaper.
I don’t know precisely where that line sits. That’s most of why I’d rather be standing close enough to see it.
The only possible opt-out
The choice was never AI or no AI. That one expired, most of us weren’t consulted about it, and there is real grief in that which I don’t want to talk anyone out of.
The choice in front of us is between AI you can see and AI you can’t. Between use you get to shape and use that happens to you.
Say no if that’s the honest answer for where your organization is right now. Some of you have good reasons and I’m not going to pretend otherwise. Just know what the no is buying.
Not an organization without AI. An organization without a say.
Anthralytic helps mission-driven, resource-constrained organizations with impact strategy, measurement, and reporting. If you’re writing an AI policy and want something more useful than a prohibition, that’s a conversation I’d like to have.
How I made this: I use AI as a writing tool. The meeting, the argument, the transcript story, and the discomfort are mine. Claude pulled some of my published work for cohesion and I built my argument and we revised and iterated together. The thinking is mine. The typing was shared.
