What "Responsible AI" Actually Asks of a Foundation

Sep 7

5 min read

TL;DR: “Responsible AI” has become a phrase foundations say to signal they are paying attention. But responsibility isn’t a statement you make — it’s a set of things you are willing to do differently: how you fund, what you ask of grantees, and who gets a say. For a foundation, responsible AI is less about the technology and more about the power you already hold.

I have been in a lot of rooms lately where a foundation leader says the words “responsible AI” with real conviction. And I believe the conviction. What I am less sure about is whether we mean the same thing.

Most of the time, “responsible AI” arrives as a value — something to affirm, like transparency or equity. It sits well in a strategy deck. But responsibility, if it means anything at all, is not a value you hold. It is a set of things you are willing to do differently. And for a foundation — an institution built on deciding where money goes and who gets trusted with it — those things are surprisingly specific. Some of them are uncomfortable. That is usually how I know they are the real ones.

So what does “responsible” actually mean here?

Responsible is a relational word. It asks who, not just what. Who bears the cost when a model gets it wrong? Who was in the room when the tool was chosen? Who was never asked at all?

When I sit with a foundation on this, I am not really asking about their technology. I am asking about their relationships — with grantees, with the communities those grantees serve, with the vendors quietly selling into the whole ecosystem. My AI Equity Framework has six layers, and none of them are about the model itself. Data equity, transparency, accountability, resource equity, co-design, and power. Every one is a question about people. The technology is the easy part. The relationships are where responsibility either happens or doesn’t.

Isn’t a good AI policy enough?

This is the honest counterargument, and I don’t want to wave it away. A written AI policy is where many thoughtful foundations start, and it is a good instinct. It forces a conversation. It slows the reflex to adopt.

But here is the gap I keep running into. A policy governs the foundation’s own behavior — how staff use these tools, what the institution will and won’t do internally. The harder question is what you fund. You can hold an immaculate internal AI policy and still write checks that pay for tools built on scraped community data, deployed on the very people your mission claims to serve. The policy protects the institution. It does not, on its own, protect the community. That gap — between what you govern and what you fund — is where responsibility actually lives.

What does responsible AI ask of how you fund?

This is the part foundations tend to skip, because it points inward. Most AI harm in the nonprofit world does not come from a nonprofit being reckless. It comes from a nonprofit being under-resourced — handed a shiny tool, no time to vet it, no budget to maintain it, and a funder who wants to see “innovation” by the next report.

So responsible funding starts before the tool. It means paying for readiness — the staff time, the data hygiene, the plain human capacity to ask hard questions — not just the software license that looks good in a grant narrative. It means asking grantees about their AI use in a way that invites honesty instead of punishing it. If your application rewards the organizations claiming the most sophistication, you will get performances of sophistication. The ones telling you the truth — “we are not ready, and we are worried” — are often being the most responsible of all.

Who answers when the model is wrong?

Picture it concretely. A model inside a grantee’s program makes a wrong call — it mis-sorts someone, flags the wrong family, quietly deprioritizes a person who needed to be first. Who answers for that?

Usually the honest answer is: no one. The vendor points to the terms of service. The nonprofit points to the vendor. The foundation points to the nonprofit’s autonomy. And the person actually harmed absorbs the whole thing in silence, because they were never a party to any of it. Foundations almost never place themselves anywhere in that accountability chain — they funded it, then stepped back. Responsible AI asks you to stay in the chain. Not to micromanage grantees, but to refuse the comfort of being the one actor always upstream of the harm and never answerable for it.

What would it actually look like to move from the word to the practice?

Here is where I try to be concrete, because “be responsible” is not an instruction anyone can follow. A handful of commitments that change what a foundation does, not just what it says:

Fund the readiness, not just the tool. Budget for the staff time, training, and data work that make any AI adoption safe — before a single license is bought.
Ask about AI on the application, and reward honesty. Make room for “we are not ready” to be a fundable answer, so grantees stop performing a confidence they don’t have.
Put community voice before procurement. Co-design means the people affected by a tool have a say before it is chosen, not a survey after it is deployed.
Name who is accountable, in writing. Decide, out loud, who answers when a model is wrong — and make sure it is not only the person with the least power.
Leave some power on the table. Self-determination means communities get to decline a tool, keep their own data, and define success on their own terms — even when your logic model would prefer otherwise.

Isn’t this asking foundations to become AI experts?

No — and I want to be careful here, because the fear of not being expert enough is what freezes a lot of good funders. You do not need to understand transformer architecture to fund responsibly. You need to understand power, because you already hold it. Technical fluency can be borrowed or hired. The willingness to be questioned by the people your money affects cannot be. That part is yours.

I don’t have this fully figured out, and I am wary of anyone who says they do. The field is moving faster than our wisdom about it. But I am fairly sure of this much: responsible AI, for a foundation, is not a thing you announce — it is the willingness to let your own power be questioned by the people it touches.

Frequently asked questions

What does responsible AI mean for a foundation specifically?

For a foundation, responsible AI is less about the tools it uses internally and more about how it funds. It means resourcing the readiness that makes AI adoption safe, asking grantees about AI in ways that reward honesty, keeping the foundation itself inside the chain of accountability when a model causes harm, and giving affected communities real say over the tools deployed on them. The technology is secondary; the power dynamics are the point.

Does a foundation need an AI policy before funding AI work?

An internal AI policy is a good starting point, but it is not sufficient on its own. A policy governs how the foundation and its staff use AI internally, while most real risk lives in what the foundation funds — tools built on questionable data or deployed on under-resourced grantees. A policy should be paired with funding practices that account for community impact, or it protects the institution without protecting the people the mission serves.

How can a foundation support responsible AI without deep technical expertise?

Technical expertise is helpful but not the prerequisite most funders assume it is. A foundation can borrow or hire technical fluency, but it cannot outsource the willingness to be accountable to the communities its funding affects. In practice that means budgeting for grantee readiness, inviting honest answers about AI use, involving affected people before tools are chosen, and naming clearly who answers when something goes wrong.