1. Open — The system you never met
Good morning. I want to start with a question, and I want you to answer it honestly, in your own head. Not out loud.
How many artificial intelligence systems made a decision about your life this week?
Not touched your life. Not recommended a video. Made a decision. Approved something. Denied something. Ranked you. Flagged you. Scored you.
Most of you don't know. That's not a failure of your attention. That's the design.
We are living through the largest, fastest deployment of automated decision-making in human history — and the public, the people those decisions are being made about, has almost no way to see it, name it, or challenge it.
That is the problem I came here to talk about. And I came with a solution.
2. The gap — Why voluntary ethics failed
For the last decade, we've been told the answer to AI harm is responsible AI. Ethics boards. Principles. Model cards. Trust and safety teams.
I believed in that once. I want to tell you why I don't anymore.
Voluntary ethics only works when three things are true: the people writing the principles have power over the people shipping the product; there is a cost to violating them; and the public can verify the claim. In commercial AI today, none of those three are reliably true.
Ethics boards get dissolved the quarter before launch. Model cards are marketing. "Responsible AI" has become a compliance function that reports to the same executive whose bonus depends on shipping the model.
And when a system fails — when it denies the benefit, misreads the tumor, flags the wrong parent, sentences the wrong kid — there is no public record it existed in the first place.
You cannot hold accountable what you cannot see.
3. The proposal — A public registry, kept in public
So we built one. It's called the Ethos AI Registry, and it is exactly what it sounds like: a public, searchable list of AI systems in deployment, with a scored audit against a real standard — the NIST AI Risk Management Framework — attached to each one.
Every entry answers three questions in plain English:
One. What does this system do, and to whom?
Two. Who is accountable when it's wrong?
Three. How would we know if it was wrong?
If a system can't answer those three questions in public, it should not be making decisions about the public. That is not a radical statement. That is the same standard we already apply to bridges, to drugs, to elevators, and to elections.
4. The proof — What it looks like when it works
I'll give you one example, and then I'll get out of your way.
When a community member files a report through our registry — a constituent says, "this system denied me and I don't understand why" — it doesn't disappear into a support queue. It becomes a public record, attached to the system that produced it, timestamped, and searchable by every journalist, every legislator, and every other citizen in the country.
One report is a complaint. Ten reports against the same system, from ten strangers who never met each other, is evidence. That is the sensor network public interest AI has never had. We are building it.
5. The ask — Come keep the receipts
I'm not here to ask you to trust me. I'm here to ask you to stop trusting anyone — including me — who won't show their work.
If you're a policymaker: mandate the registry, or one like it, in your jurisdiction.
If you're a journalist: use it. Cite it. Break stories out of it.
If you're a civic technologist, a founder, a public servant, a teacher, a parent — file a report the next time a system gets it wrong. That is how public infrastructure gets built. One receipt at a time.
The systems making decisions about us already exist. The only question left is whether the record of them exists too.
Let's go keep the receipts. Thank you.
