Two things happened this week. They looked like different stories.

On June 25, the White House asked OpenAI to limit the release of its next model, GPT-5.6, to a small group of government-approved partners. The request came from the Office of the National Cyber Director and the Office of Science and Technology Policy. Commerce Secretary Howard Lutnick separately advised OpenAI’s CEO not to lauch without cross-agency sign-off. OpenAI complied—not because it was legally required to, but because, as Sam Altman told employees, there was simply no better option in a “strange moment” with no true regulatory structure in place.¹

Around the same time, Wired publisehd a profile of Anthropic under a headline that named what the company’s critics had been saying for months: that Anthropic believes its own success is the key to making AI safe. Anthropic’s answer to that was, in essence, yes — and that this is precisely the point.²

One story was about a government reaching for a lever it does not yet own. The other was about a company claiming to be the one who should hold it. Together, they describe the same problem: when a technology becomes powerful enough that everyone agrees it needs to be governed, and no legitimate governing structure exists, the question of who fills that role does not stay open for long.

The intervention in OpenAI’s launch was unprecedented — and instructive in ways that go beyond the precedent.

GPT-5.6, according to people familiar with the situation, has capabilities comparable to Anthropic’s Mythos model: sophisticated enough in cybersecurity applications that Washington became alarmed.³ The concern was real. Two weeks earlier, the Commerce Department had ordered Anthropic to cut off all foreign access to its most advanced models for the same class of reasons. The government’s instinct to pump the brakes on GPT-5.6 was not invented.

But instinct is not policy. President Trump signed an executive order earlier this month that called on AI companies to voluntarily submit advanced models for government review thirty days before release. The framework for that process has not been built. The debates within the administration over how restrictive it should be delayed the order for weeks.⁴ So when GPT-5.6 approached readiness, what existed was not a system. What existed was Commerce Secretary Lutnick’s phone number, and the judgment of officials who had watched what happened with Anthropic and did not want a repeat.

A phone call produced a compliance. A source quoted by CNN described the overall approach as “ad hoc, personalized, opaque, possibly lawless.”⁵ That description is accurate. It is also, for now, the entirety of what AI safety governance looks like at the level of the most powerful models in the world.

This matters not because the outcome was wrong — a brief delay while the government gets its bearings may be entirely reasonable — but because an outcome that happens to be right, produced by a process that has no rules, offers no protection against an outcome that is wrong. The value of a governance structure is not in the decision it produces when everyone is acting in good faith. It is in what it can do when someone is not.

Anthropic’s position is harder to argue with, which makes it more important to examine carefully.

The company does not say: trust us. It says something more structurally interesting. In a recent Wired profile, Helen Toner — a former OpenAI board member and director of Georgetown’s Center for Security and Emerging Technology — offered the clearest account of Anthropic’s logic: powerful AI is like a forest full of treasures and mosters, and all the villagers are rushing in regardless. Given that, Anthropic’s strategy is to go further into the forest than anyone else, while investing more heavily than anyone else in understanding what lives there. “People are going in the forest anyway, we have to do it first,” Toner said. “It’s just a weird enough strategy that people have a hard time hearing it.

CEO Dario Amodei has framed the same logic in terms of physics: “If you can do that, the gravitational pull you exert is so great.”

The argument has genuine force. It is not a marketing position dressed up as ethics. It reflects something real about how technological development works: that the norms governing a technology tend to be set by those closest to its frontier, and that being absent from the frontier means ceding that norm-setting to someone else. If the choice is between Anthropic at the frontier and a less careful actor at the frontier, Anthropic at the frontier may well be preferable.

But the argument also forecloses something. Once you accept that Anthropics’s presence at the frontier is itself a safety measure, every question about the company’s growth — its accumulation of capital, compute, political access, market share — becomes answerable with the same sentece: this is the price of the mission. The framework makes itself immune to the criticism it most needs to hear. Not through dishonesty, but through its own internal coherence.

There is a further complication. In February, Anthropic revised its Responsible Scaling Policy, removing the commitment it had previously made to never train a model unless it could guarantee in advance that adequate safety measures were in place.⁶ The company’s Chief Science Officer, Jared Kaplan, explained the reasoning to TIME: “We didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments … if competitors are blazing ahead.”

Read that sentence carefully. The company that claims its own growth is the condition of safe AI revised its safety commitments because its competitors were growing. The standard moved when the race moved. That is not refutation of Anthropic’s mission. But it is a demonstration of the mechanism by which a self-authored safety standard behaves under pressure — and pressure is precisely the condition under which safety standards most need to hold.

The problem, in the end, is not that Anthropic is wrong. The problem is that architecture of its argument makes it structurally difficult to determine whether it is right. When the institution claiming to protect the public also defines what protection means, the public has no independent purchase on the claim.

What both stories share is a vacancy.

When technology has historically disrupted societies faster than those societies could absorb the disruption, the response has never come from the disruptors themselves. It has come from those who bore the cost: workers who organized, communities that demanded legislation, publics who elected representatives willing to act. The process was slow, contested, and deeply imperfect. But it had a defining feature: it was authored by those with something to lose, not by those with something to gain.

That author is missing from both of this week’s stories. The White House acted through informal pressure rather than law. Anthropic governs through internal policy rather than external accountability. Both are doing what they can in the absence of something neither of them can provide: a structure that is independent of the entities it is supposed to govern.

There is a version of this argument that ends in pessimism: the institutions don’t exist, they will take decades to build, and in the meantime the technology will have already shaped the world in ways that make the institutions beside the point. I do not think that is the right conclusion. Institutions have been built under pressure before, often faster than anyone expected, when the alternative became untenable.

But they are not built by the entities whose power they are meant to constrain. And they are not built by default.

The brake exists. Two hands are on it. Neither of them was delegated to hold it, and neither cam be removed if they hold it wrong.

That is the problem this week’s news descirbed. Neither story named it. Both stories were about it.

Polanyi is an independent index of what technology does to us. This is the third letter.