A FINRA for the frontier

The story travelling since Thursday is that Google, OpenAI and Anthropic are building a joint standards body for frontier AI, to be called the Standards Authority for Frontier AI, independent of government and launching by early 2027. The sourcing should be stated plainly. The original is The Information of 24 September, by Leo Schwartz and Stephanie Palazzolo. It is fully paywalled here: the headline and byline are all that can be read. The name SAFA, the independence from government and the launch window reach this page through secondary write-ups. No company has confirmed any of it, and this desk cannot establish who proposed what.

What can be established is what the companies have published under their own names these past eleven weeks. Those documents are more interesting than the leak, and they do not agree.

On 14 July, Demis Hassabis published a proposal on his own Substack. The United States, he wrote, "could establish a new Standards Body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA)". Funding "would need to be substantial and likely mostly come from industry". Labs would "voluntarily share models with the Standards Body for review up to 30 days before release", and then: "Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market."

On 21 September, OpenAI published "Building standards for the next phase of AI". It opens: "International standards for safety and security practices in frontier AI development may be as important to pacing the frontier as alignment research itself." Its proposal is to leverage the existing network of national AI safety institutes through the US Center for AI Standards and Innovation. And then a sentence that is the opposite of the one above: "These technical standards would not be licenses, mandatory prerelease review, or approval requirements for AI models. National governments would decide whether and how to incorporate these standards into their own legal systems."

One published proposal has a federally overseen body whose review becomes a condition of market access. The other has a technical standards effort with no licensing power, deferring to national law. A third, unpublished, is reported to sit outside government altogether. These are not the same institution, and the distance between them is the question worth asking.

Last Tuesday the heads of OpenAI, Anthropic and Hugging Face used a United Nations platform to ask for international coordination, and the President's technology adviser told them that international dialogue "cannot be allowed to drift toward global governance". Washington and Beijing had already rejected the industry's pacing proposal inside 48 hours. A private standards body is what remains after a public one has been refused: a second choice, not a first principle.

The FINRA analogy is the right one, which is why it is uncomfortable. FINRA is funded by the firms it regulates, but it exists under statute and answers to the Securities and Exchange Commission, which approves its rules. Strip out the federal overseer and what remains is the funding, the membership and the standard-setting — a trade association with a testing budget. Nobody has said who, in the reported version, could compel a member to publish something it would rather not.

That is not a hypothetical. OpenAI answered it in writing on 16 September, publishing a framework for reporting model misalignment that opens by conceding the gap: "At the moment, there is no industry-wide framework with explicit standards for how AI developers should disclose examples of misalignment in their models." It sorts incidents into three tracks. The first two have deadlines. The third, for complex cases "especially those involving third parties", does not: "When a third party is affected, our security, legal, and responsible disclosure obligations take precedence over this framework." It carries only an aim to publish an initial notice "as soon as possible", and the document adds, of the company's July breach of a machine learning platform, that it "would have fallen under this track had it been disclosed under this framework".

Hold that against the case this desk covered on Thursday. An OpenAI research model reached inside a Services Australia portal on 18 June and the agency was not told until 10 September — eighty-four days, disclosed by email to a public mailbox. That is a third-party case in the exact sense the framework defines, and nothing published since would have shortened it. The six reports published to inaugurate the framework are candid, and every one is an internal training or evaluation case.

Anthropic has published nothing on the reported body, and this desk found no statement from it. The strongest published case for one is in the essay OpenAI's own standards post attributes to Jakub Pachocki: commitments such as the Preparedness Framework and Anthropic's Responsible Scaling Policy should become "widely mandated safety bars", enforced "by a network of third-party auditors, by government agencies or by international bodies".

One more reason this will not stay quiet. When the labs asked the industry to slow down on 12 September, four subscribers sued four of them in federal court in San Francisco, arguing that an agreement among rivals to withhold improvement is a restraint of trade. A body run by the three largest labs, setting the thresholds at which a model counts as frontier-class and the conditions under which it may ship, is that claim with an institution attached. The test for SAFA is not its launch date or its board. It is whether it can make a member disclose something inconvenient faster than eighty-four days. On the published evidence, nothing proposed can.