The Companies Aren't Avoiding Regulation. They're Designing It.

Demis Hassabis is proposing that Big AI buy itself the word “Frontier.”

I want to be fair to the narrow version of this, because the narrow version is genuinely embarrassing to argue against. Yes, model evaluation is a real and young field. Yes, the labs he is describing do fund serious safety research, and some of the people who would staff his body would be engineers of real judgment. The problem is not the worry, which I take to be real. The problem is the architecture of the proposal, which has been engineered so that the worry can never bind on anything.

Hassabis spent his final weeks as Google DeepMind CEO building the safety regulator — for his own industry. He met with Treasury Secretary Scott Bessent and White House tech adviser Michael Kratsios. He discussed the plan with rival AI labs. He published a July essay on X laying out the architecture: an industry-funded “Standards Body” modeled on the Financial Industry Regulatory Authority — the current name of the brokerage industry’s self-regulatory organization, whose lineage runs back to the National Association of Securities Dealers, organized in 1939 to keep Congress from imposing something stricter.

The parallel is an old play. After the 1929 crash, Wall Street proposed self-regulation to forestall New Deal enforcement. The result was an industry-funded body that writes and enforces its own rules, examines its own brokers, and answers, in the final instance, to the industry that pays its budget. The structure gave Congress enough visible oversight to forestall stronger action for decades. It took the 2008 financial crisis to demonstrate what a system of self-policing with industry-designed guardrails produces when the thing being policed is genuinely dangerous and the designers have a documented financial interest in the outcome.

Then he reached for a stronger analogy, and the two analogies tripped over each other. In discussions with peers, Hassabis compared the proposed body to the International Atomic Energy Agency — the intergovernmental watchdog that promotes nuclear cooperation under treaty obligations, conducts independent inspections with legal access to facilities, and can refer noncompliance to the UN Security Council. The proposed AI body would be funded by the companies it oversees, set standards through a process those companies shape, and issue a “Frontier-class” designation based on testing conducted jointly with the labs whose models are being tested. Those are category errors marching in opposite directions. The IAEA works because it sits outside the industry it inspects and carries treaty-backed power to compel access. An industry-funded body funded by the firms it evaluates is a different kind of institution entirely, and the resemblance is in the branding, not the architecture. The vocabulary of nuclear governance is borrowed authority applied to an entity with none of its structural independence, legal mandate, or enforcement capacity.

The scheme becomes less appetizing when you read the passage about how the certification would actually work. Federal agencies and the Energy Department’s national laboratories would work with AI companies to test models, and the body would establish qualifications that would label a model “Frontier-class” if it met certain standards. A label of that sort creates a currency of legitimacy for whichever interest can afford the audit, and the companies buying the label are also the ones funding the standards organization. The obvious risk in any certification regime built this way — where the industry writes the rubric, the industry pays the graders, and the graders are answerable to the industry’s executives — is not that it fails. It is that it succeeds exactly as designed. That “Frontier-class” sticker is not a test result. It is a permission slip priced as a safety certificate, and it works precisely because it is hardest on the people it purports to protect.

The leadership transition tells the same story. Google named Koray Kavukcuoglu, Hassabis’s deputy, to replace him as DeepMind CEO. Jeff Dean, the company’s longtime chief scientist and a co-founder of DeepMind, stepped down to start an AI-research startup. Alphabet CEO Sundar Pichai wrote that he and Hassabis “have been long discussing a role that allows him to put his full attention on actively shaping the future of AGI.” The company is restructuring its leadership around the proposition that designing the governance framework is more important than running the research unit — because whoever designs the governance framework determines what the rules will be before the public or the legislature has a say.

Hassabis is not alone in running this play. OpenAI’s Sam Altman proposed “a U.S.-led international forum” for AI standards — a structure that would place American labs, specifically his, at the center of global rulemaking. Anthropic endorsed a global agreement to slow development and verify compliance — a framework that would grant the existing frontier labs a role as auditors of their own competitors’ work. In June, at the G7 summit in France, Hassabis pushed the case for U.S.-led global coordination on AI standards — with Altman, Amodei, and the president in the same room. A month later, Beijing launched its own China-led AI body. Every proposal shares a common feature: each positions the company advancing it as the responsible party that should be trusted with the architecture of oversight. The two international architectures differ in flag and forum; they agree in keeping the public out of the room where the standards are set.

The administration has kept model evaluations for the most powerful tools voluntary, which is the precondition for everything that follows: if the government will not mandate independent testing, the companies will propose testing they design and fund, and the proposal will arrive pre-formatted as responsibility. The Treasury Secretary, who has warned that Anthropic’s Mythos poses a cyber threat to the financial system, is in the room. The White House tech adviser is in the room. The companies are in the room. The public is not.

Here is the exchange that should give any honest skeptic pause. In the same weeks he was proposing this body, Hassabis was telling the world, in the memo announcing his departure from operational control, that he has “been working towards AGI [his] whole life” and that “now… I feel it is close at hand.” That is a statement with two parts, and they do not resolve the way his supporters want. Either he means it — in which case the last decade has produced genuinely alarming technology and he wants the industry that benefited from it to write the compliance rules, to be the evaluator of its own evaluation — or he doesn’t, in which case the entity he is building is not a safety mechanism but a marketing operation designed to dress up the models he is about to sell as something other than what they are. The AGI-timeline claim is doing all the work of the other claim, and it is doing that work in a way that makes the belief unfalsifiable.

The substance underneath the organizational-chart maneuvering is straightforward. AI companies want to define what counts as safe. They want that definition to center their own models as the baseline. They want the testing to be conducted in ways they participate in designing. And they want the result to carry the authority of a “standards body” that the public will read as independent. The FINRA model serves this purpose precisely because it looks independent to people who do not read the funding structure — and the funding structure is the entire point. The history of industry self-regulation in technology is the history of the regulated writing the rules that regulate them, then pointing to the existence of rules as evidence that regulation has occurred.

There is a version of this story in which an industry-funded standards body does real work: establishing minimum testing protocols, creating transparency requirements, building a shared technical vocabulary for evaluating model behavior. That version would require independent funding sources, public representation on the governing board, mandatory participation, and enforcement mechanisms with legal teeth — actual failure consequences: licenses to operate revoked, models pulled from the market, criminal liability for senior executives, the kinds of things we reserve for industries we actually consider dangerous. None of those features appear in Hassabis’s proposal. What appears is the FINRA template: industry money, industry participation, industry-designed standards, and a label the companies would need to qualify for, awarded by a body the companies would fund.

What is happening now, with the administration having kept model evaluations voluntary and the testing framework’s details kept mostly private, is what every captured industry discovers at the end of a long deregulatory shell game: the option to “self-fund” becomes the option to set your own terms. That’s a lobbying project. People with engineering talent can be honest, and even people with a financial stake can be confused about the source of their own good intentions. The companies aren’t refusing regulation. They’re designing it themselves, and the one thing the plan cannot survive is independence. Independence is the only thing it would cost them.