AI's big three are building a safety regulator. They'd also be the ones it regulates.

The plan: a FINRA for frontier AI

The three companies that build the most powerful AI models want to set the safety rules for the most powerful AI models. According to The Information, reported this week by CIO and others, OpenAI, Google and Anthropic are preparing to found the Standards Authority for Frontier AI, or SAFA, an industry body that would set guidelines around risk assessment, testing and pre-release review for frontier models, with a target launch as early as 2027. It is reported and planned, not launched, and no charter has been published.

The template is explicit and revealing: FINRA, the Financial Industry Regulatory Authority. FINRA writes and enforces rules for US brokerages, is funded by the industry it polices, and operates as a self-regulatory organization under the supervision of the Securities and Exchange Commission. Copy that design onto AI and you get a body that is private, industry-funded, and, in the best version, backstopped by a government supervisor. The whole story turns on whether that backstop is real or decorative.

How we got here: from a White House order to self-rule

The idea has a clear author. On 14 July, Google DeepMind chief executive Demis Hassabis published an essay, "A Framework for Frontier AI and the Dawning of a New Age," calling for a US-led, FINRA-style standards body. Under his proposal, labs would voluntarily share models with the body up to 30 days before release for safety testing, and once that regime proved itself, passing it could become mandatory to deploy a model in the US market. He told Axios he wanted it running before year end.

The important detail is what changed between July and now. The companies initially pursued a public-private partnership under federal oversight. After a draft White House executive order was put on hold, they shifted toward industry-run self-regulation instead. That is a meaningful downgrade in who holds the leash, from a government that convenes the standards to an industry that convenes itself, and it happened not because self-rule is better but because the government route stalled.

What "FINRA for AI" actually requires

FINRA is not a fig leaf, and neither, in its serious proposals, is this. The most rigorous public blueprint, a Lawfare analysis by the Institute for Progress's Mark Thomas, describes a supervised self-regulatory organization with real teeth: mandatory membership for anyone training models above a compute threshold, a government supervisor with authority to approve the rules and confirm board members, and a board where independent governors outnumber industry. That version could work, because the government keeps a veto.

The failure mode is the version without the supervisor. Strip out the binding government backstop and "self-regulatory body" becomes a trade association that grades its own members' homework and calls the grades standards. The gap between those two outcomes is the entire question, and nothing reported about SAFA yet tells you which one it will be.

The conflict sitting at the center

Here is the tension that no amount of credentialed board members erases: the organization would be founded and funded by the three firms whose models it regulates. OpenAI's own 21 September standards paper argues, without apparent irony, that its "rationale for standards is rooted in avoiding the concentration of power." A rulebook written and paid for by the three most powerful labs is an odd instrument for de-concentrating power, and smaller and open-source developers have a specific fear: that membership thresholds and audit requirements calibrated by incumbents will function as a moat, compliance the big labs can absorb and startups cannot.

There is a real argument on the other side. Standards genuinely can give outsiders a say and create shared definitions of what safe looks like, which is better than every lab marking its own exam in private. OpenAI's paper makes exactly that case, and it is not wrong. But "better than nothing, written by the incumbents" is a low bar dressed up as governance.

Why now: writing the rules before Washington does

The timing is not subtle. OpenAI published its standards paper on 21 September as heads of state gathered for the UN General Assembly, and the SAFA reporting followed days later. The strategic logic is first-mover advantage on the rulebook: whoever drafts the standards shapes them around their own practices, and doing it now, while Washington's own AI order is stalled and the EU has already written binding AI law with dated deadlines, lets the labs define the terms before a government does it for them. That is smart. It is also exactly why the rest of us should read the fine print before applauding.

Our take

An industry-funded safety regulator for frontier AI is not automatically a bad idea, and FINRA proves a supervised version can work. But the version worth having is the one with a government supervisor holding a real veto, and the version being reported is the one the labs reached for after the government route stalled. Judge SAFA on one question when its charter appears: who can overrule the three companies that pay for it? If the answer is a government body with teeth, this is progress. If the answer is no one, it is the most powerful companies in the field writing their own permission slips, and calling it safety.