The case for accountable AI infrastructure

I have been writing and saying the same argument for a year and a half now, and I want to try once more, in full, because it is the only argument that matters for Curos. It goes like this:

Artificial intelligence is becoming infrastructure. Infrastructure must answer to the public, not to the short-term interests of its owners. Therefore AI should be built, at least in part, by organizations that have no owners — non-profits, held accountable in public.

The first premise is not controversial anymore. Ask anyone whether AI is becoming as foundational as the power grid or the road system, and most people say yes. The conclusion is the part that is hard, and the hard part is not the idea — it is the thousands of unglamorous decisions that stand between an idea and an infrastructure that actually answers to people.

What accountability means in practice

A year and a half in, I can say what accountable means at Curos, because we have had to make it concrete:

Money you can check. Every year we publish audited numbers: where funding comes from, where it goes, what we refused. Concentration is the risk — no single funder above 30%, and none with a product veto. You can read the numbers, and you can disagree with them.

Testing you can inspect. Every model release is gated by an evaluation suite, reviewed by a committee with the power to delay, and published with a report that lists what we did not test. You can read the report. When we were wrong, we wrote it down.

A structure you can hold. The charter fixes the purpose. The board is majority-independent. The community elects a seat. None of this makes us good — structure never guarantees virtue — but it makes failure visible, and visible failure is fixable.

What free means

Ichnus is free for the same reason the first premise matters: if AI is infrastructure, it should not be rationed by ability to pay. Free is a decision about who the technology serves, and it is the decision that most concentrates the mind. A product with no revenue from its users has no incentive to exploit them, and that alignment is worth more than any mission statement.

The cost is real. Free means we cannot buy our way out of hard questions. Our compute budget is a fraction of what for-profit labs spend. But the constraint is the identity: we are the AI that answers to its users, because structurally we cannot answer to anyone else.

What we do not claim

I want to be as honest about the limits as about the argument.

A non-profit structure does not make us wise. Our models have limitations, our evaluations have gaps, and our judgment has been wrong before and will be again. The point of the structure is not to make us right — it is to make us accountable, which is the next best thing.

And the broader system will not fix itself. One non-profit does not offset the consolidation of the most important technology of the century into a few balance sheets. Curos is an existence proof that the alternative is buildable. It is not, by itself, the alternative.

What we are asking

The community that uses Ichnus has a standing role in how it behaves: the council, the public red-teaming, the board seat, and the simple act of telling us where the product is wrong. That is not a consulting exercise; it is the accountability mechanism, and it only works if people use it.

Infrastructure is what people keep — and what people keep is what they can see, test, and change. We are trying to be that kind of infrastructure. Check us, hold us to it, and if a real organization could do it better, build that organization. That is the entire ask.

All news →