Last week, Anthropic CEO Dario Amodei published an essay arguing the AI industry needs to slow down. Within hours, OpenAI's Sam Altman agreed. Both said the same thing in almost the same words: AI companies need to coordinate on safety standards before capability outruns control.

To actually do that legally, they're asking the U.S. government for an antitrust waiver. That's the detail that should stop you.

Antitrust law exists for exactly one reason: to stop competitors from getting in a room together and deciding the rules for everyone else. That's not a technicality. That's the whole point of the law. When the two most powerful AI labs in the world ask for permission to have that conversation anyway, they're telling you, out loud, that what they want to do would normally be illegal.

Maybe it should be illegal. Let me tell you why I think there's a better way.

I spent years inside one of the most regulated industries in the country: debt collection.

Before the Fair Debt Collection Practices Act, it was the Wild West. I heard the stories, collectors calling people pretending to be from a law office, working the fear until the debt got paid, whatever else it took. Some of it was flatly illegal even then. A lot of it lived in the gray area, and gray areas get pushed until someone draws a line.

Once the regulation landed, the industry didn't become perfect. It never does. But every company had the same rulebook. Nobody could out-compete an honest business by being willing to lie harder. The line got drawn once, publicly, by someone with no stake in who won.

That's what AI is missing right now.

Every AI company is in the same trap collectors were in before regulation: pushing into gray areas not because they think it's right, but because if they don't, a competitor will, and they'll fall behind. Amodei's essay is basically an admission of that trap. His fix, get permission to coordinate privately, just moves the trap somewhere with less sunlight.

We've solved this exact shape of problem before, in a different industry too. After Enron, nobody trusted that a public company's own hand-picked, self-paid auditor would tell the truth about its books. So Sarbanes-Oxley created an independent board to oversee auditors instead of leaving companies to pay their own referees. It worked because the people checking the work didn't answer to the people being checked.

Here's what I'd build instead of a waiver: a federal agency modeled on that same fix.

Call it the Federal AI Practices Bureau. It sets the safety standards Amodei is asking labs to agree on privately, except publicly, where anyone can see how the standards got set. It hires the third-party evaluators, the METRs of the world, and puts those contracts out to competitive bid, the same way any government contract works. The AI companies pay into it, the same way regulated industries fund their own oversight today. Nobody's paying their own grader anymore, and nobody needs an antitrust exemption to talk to a competitor, because they're not setting the rules together. A regulator is setting them for everyone.

Is this slower than an essay and a handshake? Yes. Building a real agency takes time, and AI capability isn't waiting around for Congress. And a domestic bureau has no authority over what gets built in Beijing. That piece runs through the State Department, not the FAIPB. But diplomacy with a rival government isn't a light switch. The closest working model we have for getting a rival nation to accept outside verification on dangerous technology is something like the International Atomic Energy Agency: decades of negotiation and inspections, and still imperfect compliance from the countries most determined to cheat.

This isn't hypothetical. The 2025 CIA assessment on COVID's origin leans toward an accidental lab leak, at low confidence, in part because nobody outside that lab had the access to know for certain either way. That's what happens when research runs without independent eyes on it. Whatever actually happened in Wuhan, the lesson holds for AI: a lab with no outside verification is a lab where accidents don't get caught until it's too late to matter.

State Department involvement opens that track. It doesn't finish it overnight, and nobody should pretend otherwise.

But a slow, public referee beats a fast, private cartel every time. We've done this twice already: once for the people collecting your debt, once for the people auditing your public companies. It's time to do it a third time, before the gray areas become the whole map.

References and Further Reading

The essay and the response

The regulatory precedents

What the third-party evaluators actually check for

The COVID origin assessment

Originally published on LinkedIn. Comments and the conversation live there.