The Labs Are Building a FINRA for AI. FINRA Needed an Act of Congress.

On September 15 the chair of the Federal Trade Commission told an audience at Georgetown that policymakers should be "deeply suspicious" of AI companies that ask for antitrust exemptions while lobbying for regulation. On the same day, OpenAI's global policy chief told reporters in Washington that OpenAI, Anthropic and Google DeepMind are coordinating on safety anyway, without a waiver.

Both of those are now true at once. The gap between them is where AI self-regulation actually lives, and it is narrower than the announcements suggest.

## The model they picked has a statute behind it

The three labs are designing an industry body modeled on FINRA, the US financial self-regulator. OpenAI confirmed the plan, which originated in a proposal from a Google official. Mandate, membership and legal authority are all still unfinalised.

That last item is not a detail to be sorted out later. It is the entire question.

FINRA is not a club that firms joined because they cared about investor protection. It operates under the Securities Exchange Act, registration with it is effectively mandatory for broker-dealers, the SEC oversees and can override its rulemaking, and it can fine, suspend and expel a member from the business. The self-regulation works because Congress put a statute underneath it and a federal regulator on top of it.

A body announced by three companies has none of that. It cannot compel membership. It cannot stop a non-member from shipping a model. And it cannot write the one rule that would matter most, an agreement among competitors about what none of them will build, because that is precisely the kind of agreement antitrust law exists to prevent.

## The waiver was the ask, and it was refused

Anthropic had called for antitrust permission to coordinate a slowdown with rivals. Chair Andrew Ferguson's public warning was the answer, and his reasoning is not unreasonable. A firm asking for both the rules and an exemption from the rules is asking for a competitive moat with a safety label on it.

So the labs kept the coordination and dropped the waiver request. That is the worst of the available combinations. Safety discussions between the three largest model developers now carry real legal exposure, which means they will stay narrow, undocumented, and carefully steered away from anything that looks like an agreement on output. You cannot build a credible regulator out of conversations that counsel will not let you write down.

## Washington is not going to supply the statute this year

The obvious fix is legislation: a narrow, supervised safe harbour covering specified safety coordination, which Congress has granted to other industries before. OpenAI's Chris Lehane said the company would back bipartisan legislation on catastrophic risk.

Then look at what such a bill has to pass through. Senator Elizabeth Warren is calling for an immediate pause on advanced development, delivered at a roundtable hosted by Bernie Sanders. The President is dismissing AI risk warnings as hoaxes. There is a pause caucus and a hoax caucus, and not much in between. That is not the configuration that produces a technical safe-harbour bill in an election-shortened session.

We argued yesterday that nobody currently has the standing to enforce a slowdown. This is the mechanism sitting underneath that conclusion. The problem is not an absence of will among the labs. It is an absence of a legal instrument, and the body being announced does not create one.

## What to do with this

Two things to watch, both cheap to check.

First, when the body is formally announced, look for a published mandate and a membership list. FINRA's authority begins with who is required to join. If the announcement names neither, read it as a trade association with a safety brief rather than a governance event, and price it accordingly.

Second, watch whether anyone drafts a narrow statutory safe harbour instead of a blanket exemption. A carve-out limited to defined categories of safety information, with an agency supervising its use, is the version that could survive Ferguson's objection. If that bill never appears, the FINRA comparison stays a metaphor, and the frontier stays governed by whatever three labs can say to each other with the lawyers in the room.