It's Not Whether You Use AI. It's Who You Back to Build It.

Vu Le's essay on the harms of AI in the social sector dropped yesterday. He named real harms. I spent the day digging into the underlying research. The strongest claims hold up.

Our sector hasn't been racing toward AI tools. The opposite. Most of us have been hesitating in the library while the music bumps in the next room. Vu's piece arrived right as some of us were finally walking toward the door. It's going to convince a lot of those people to sit back down.

Underneath the post is a different conversation. In the comments. In the DMs.

The harms are real. Refusing won't fix them. Using carefully won't fix them. Government isn't coming in time. So what am I supposed to do?

I directed $700 million in philanthropy at Google.org. Last September I was on the other side, pulling retirement money to keep what I'm building with AI alive. Both seats taught me the same thing about where our power actually lives.

It's not in the answer we keep grabbing for.

We're asking the wrong question

The sector has been stuck on whether we use AI.

Frame it that way, and we're consumers. Opt in. Opt out. Use it carefully. The companies that build these models will ship the next one whether we participate or not.

Our power is larger than that. What gets built, by whom, for whom, accountable to which mission, is still being decided. We help decide it. With what we fund. With who we name. With who we back.

Whether to use it is the wrong question. The right question is who we back to build the thing that comes after.

What Vu got right

His strongest claim is one we should read and remember.

Last year, researchers at Stanford and the University of Chicago fed leading AI models the same questions written two ways. Once in standard English. Once in Black English. Nothing else changed.

The models recommended more criminal convictions when asked to judge the speakers of Black English. More death sentences. Lower-status jobs. The researchers then trained the models to be more polite. The models stopped saying racist things out loud. They kept making the same decisions.

That's a peer-reviewed result. If you build, fund, or deploy AI in this sector, you need to take heed. "The polite tone these models have been taught isn't evidence the bias is gone."

Vu's other strongest claims rest on similar evidence. Court records. Peer-reviewed surveys. FEC filings.

The workers in Nairobi paid pennies to watch the worst content humans have ever made.

The surveillance. The political alignments of certain CEOs.

This isn't moral panic. It's the paper trail.

What Vu leaves out

He also leaves things out. They matter. They're the case for what's possible.

A purpose-built therapy chatbot, designed with clinicians, cut depression symptoms by half in a randomized trial.

AI cuts professional writing time by forty percent. It does the most for the workers with the least training, not the experts.

Blind, deaf, dyslexic, and motor-disabled users have built workflows on these tools that didn't exist three years ago.

None of this refutes Vu. It shows a different thing. AI built for the work, by people accountable to the work, can do what general-purpose AI can't. Refusing all of it gives up that ground.

That's the case for backing builders.

The three answers we reach for

Most of us reach for one of three answers. Each feels like agency, but none of them changes what gets built.

Refuse it. A personal stand has its own integrity. Bearing witness without complicity matters. Witness alone doesn't build something different. The companies don't need your usage to ship the next model.

Use it carefully. Necessary. Not enough. Careful use makes a single workflow a little better. It doesn't change who decides what gets built next.

Wait for government. Keep pressing, especially at the state and city level. Congress is locked. The industry has put more than a hundred million dollars into political efforts designed to weaken or preempt state AI safety laws. The federal action we need isn't coming in time. Policy matters. Waiting isn't a strategy.

So the question is real. What's left?

Where our power actually lives

What's left is the people close to the work.

The AI architecture is being built by the frontier labs. The application, what tools serve what mission, is still ours.

There's a layer of AI being built right now by people who came up inside this sector. Some are running pilots inside nonprofits. Some are bootstrapping on personal savings. Some are running open-source experiments out of university labs. Most are invisible.

That's the layer we can change.

Refusing doesn't build it. Using carefully doesn't build it. Waiting doesn't build it.

People we back build it.

Three things to do right now

Three concrete moves. The dollars aren't equivalent, but the importance is.

1) If you sit in a funding seat. Write at least one check this year to an organization building something socially impactful with AI. A hundred thousand to a quarter million dollars, structured so success recycles the capital and failure still counts as a mission-aligned grant. The instruments exist. Your foundation counsel knows them.

Foundations move a hundred and ten billion dollars a year. About a penny on the dollar comes back as recoverable capital. Almost none of that penny reaches the kind of AI being built by startups. What's missing is the practice.

Before you write the check, hand them Vu's essay. Ask which harms their work makes worse, which ones it doesn't, and how they know. A builder who flinches at that question isn't ready to build for our sector.

(The longer version of this argument is in I Gave Away $700 Million. Then I Tried to Raise $350,000. This is the structural ask underneath it.)

2) If you lead inside a nonprofit. Pick one workflow your team would build differently if you owned the tool. Name it. Tell a funder what it is. The next time someone asks how your team is using AI, flip the question. "We know the tool our work needs. Will you help us build it, test it, or buy it from someone whose mission lines up with ours?"

3) If neither of those describes you. Make a builder visible. Most organizations working on this die for one reason. Nobody can find them.

Name one of these orgs publicly in the next thirty days. Hire them for a small piece of work. Subscribe to what they ship. Send their name to a funder you know. Repost what they write. Visibility is its own capital. It's also the move that doesn't require a checkbook.

Enough of these, by enough people, changes what becomes normal.

What this won't do

Backing builders doesn't stop the frontier labs from shipping what they're shipping.

It doesn't replace policy. We still need state attorneys general. State legislatures. Federal action when it finally comes. Keep the pressure on.

It doesn't make the harms disappear.

And it doesn't put the public in charge of AI.

That's the deepest critique of this argument. The people deciding what gets built are still a small group choosing for the everyone else. Funders. Founders. Networked sector folks. The asking is happening in the rooms we already sit in, not in the rooms where the cost is being paid.

A question that belongs to the public doesn't have an honorable private answer. The fight for AI that's democratically governed, owned and steered by the people whose lives it's reshaping, is the bigger one. It's also the right one.

Public ownership stakes on AI infrastructure are on the table for anyone with the nerve to put them there. That work matters. We should be loud about it.

It's also a multi-decade project, absent a catastrophe that shifts public opinion fast. In the meantime, the question is what we do with the agency we already have.

Philanthropy moves more than enough capital to measurably move the needle on what gets built in the next four quarters. And the four after that. And the four after that.

Not as a substitute for democratic governance. As the work of the meantime.

What it does is seed a different ecosystem. Tools built by people who understand the work. Accountable to missions that aren't solely focused on shareholder returns. It moves the ground under our feet, even a little, toward the kind of AI this sector would actually trust.

That is not nothing.

What's left to do

The next time a colleague forwards Vu's essay and asks what to do, the honest answer is shorter than the conversation we've been having here.

The music isn't stopping. The architecture is being built. The next model is shipping whether we participate or not.

Refusing won't change that. Using carefully won't change that. Waiting won't change that.

The people who will change it are already in your network. Most of them aren't visible enough yet.

Find one. Back them. Tell the rest of us who they are.

Who's building the AI this sector would actually trust?

I'll start. We built the Kindora Claude Connector last month. It lets people search more than a hundred thousand open grants from inside any AI assistant. We gave it away free.

People keep asking why we didn't charge for it. The honest answer is that's what AI built for our work should do. Save people time they can reinvest in relationships. Make their work stronger. Delight them. We believe generosity will come back one way or another.

One example. Not the only one. Not even the best one. The one I know best.

Share yours.

Justin Steele is co-founder and CEO of Kindora, a Public Benefit Corporation using AI to help social-sector organizations find and win funding. He previously directed nearly $700 million in philanthropy at Google.org over a decade and serves as a trustee at The San Francisco Foundation. This piece was originally published on LinkedIn.