Anthropic is going public on the claim that its AI might kill you. The S‑1 will talk about market opportunity and technical moats and responsible scaling, but the logic is inescapable. A company that warns the world its products pose an existential threat to human civilization is making a claim about the value of what it controls. If the models are as dangerous as Anthropic says they are, then the only company positioned to build them responsibly — the only company that pauses, that shuts down, that warns — is worth whatever it asks. The threat narrative is the moat.
It is true that Stuart Russell is a serious computer scientist and not an Anthropic spokesperson, and it is true that the company has been more forthright than some of its competitors about the dangers of recursive self‑improvement and the potential for catastrophic outcomes. The trouble is that the door Anthropic wants locked is the one that would keep other firms out of the very market it now stands to dominate. What the company presents as a plea for a global AI pause is, on examination, a monopoly‑enforcement demand dressed in the language of existential risk.
Consider the sequence of events in June. On the 5th, Anthropic urged the industry to coordinate a pause on advanced AI development. On the 9th, it released Fable 5 to the general public, with guardrails, under the product name Claude. Three days later, the White House issued an export‑control directive banning all foreign nationals from accessing Fable 5 and Mythos 5, a directive that immediately cut off many of the company’s own researchers. Anthropic responded by shutting the models down entirely. The company that called for a pause had already released the product. The government that restricted access was restricting access to a system that had been public for seventy‑two hours. The net effect is not an isolated incident but the latest turn in a cycle that simultaneously inflates the perceived threat and positions Anthropic as the only entity capable of containing it.
If you are genuinely in the grip of a runaway feedback loop that threatens human extinction, you do not first offer the system to the general public through a product called Claude. You stop the centrifuges. One leading AI CEO told Russell he did not expect serious regulation until there was a “Chernobyl‑scale disaster.” That is not the statement of a safety‑first actor. It is the statement of someone who has understood that safety, for his firm, is a bargaining chip.
The recursive‑self‑improvement claim itself deserves the same scrutiny. No one outside Anthropic can audit Mythos 5’s cyberattack capabilities or verify the early signs of RSI — the technical term for a system improving its own intelligence in a runaway feedback loop — that the company’s blog post describes. The claims are unverifiable by the public, unfalsifiable in the short term, and enormously valuable to the company making them. Cory Doctorow, naming what the historian Lee Vinsel calls “criti‑hype,” observes that AI critics are “prone to engaging in criti‑hype: criticizing something by repeating its boosters’ claims without interrogating them to see if they’re true.” Russell’s column is a clean example. He takes Anthropic’s RSI claim at face value. He takes the cyberattack capability at face value. He takes the “one in six chance of human extinction” — attributed to unnamed CEOs — and builds a regulatory argument on top of it. The argument may be correct. But the evidentiary foundation is the company’s own press releases, given weight by a distinguished scientist’s engagement.
Adam Becker, in More Everything Forever, describes how the ideology of technological salvation “channels attention into abstract future possibilities, thereby avoiding politically uncomfortable questions about present‑day redistribution and power structures.” The existential‑risk framing does exactly this work. While the discourse focuses on whether Anthropic’s models might someday end human civilization, the concrete questions go unasked: Who owns the training data? Who paid for the compute? What happens to the researchers whose labour built the systems now replacing them? What does it mean that a company’s market position depends on its products being perceived as catastrophically dangerous?
Doctorow and Giblin taught us to recognise chokepoint capitalism when we see it: a firm positions itself between a necessary resource and everyone who needs it, then extracts rent on every transaction. General‑purpose frontier AI, if it proves to be as powerful as its makers claim, will be the ultimate chokepoint. Owning the model that everyone must license is not a technology business; it is a tollbooth. And the tollbooth operator always, eventually, enshittifies — first generous to users, then to business customers, then to shareholders alone. Anthropic may be a more careful steward than some of its peers, but the four forces that constrain enshittification are already being dismantled around it. Competition is being replaced by regulatory barriers. Regulation is being shaped by the regulated. Adversarial interoperability is criminalised by the very same export‑control and anti‑circumvention architecture. And the scarce, mission‑driven engineers who might once have refused to build something dangerous are being replaced by the same models that now write all their code. A company whose leading researchers “no longer write any code at all,” as Russell reports, has already turned its own workers into reverse‑centaurs — human peripherals to a machine that is running faster than they can supervise.
The White House’s export ban functions as a state‑built moat. It locks foreign competitors and foreign talent out of the leading edge while leaving Anthropic free to pursue its IPO and its next‑generation systems internally — security‑cleared American personnel only. Cybersecurity experts have already urged the White House to reverse the restrictions, but the damage they might do to the open research environment is, in a sense, the point. The company’s decision to shut the models down entirely, far from being an act of civic conscience, looks like a negotiating tactic: give us a regulatory settlement that preserves our lead, or the engines stay off.
The licensing regime Russell calls for is the endgame. A minimum safety standard before a system can be built — “how we handle nuclear power, airplanes, buildings, elevators, hairdressers and sandwich makers,” he writes. The analogy is appealing. It is also structurally identical to every regulatory framework that incumbent industries have used to lock out smaller competitors since the invention of professional licensing. Nuclear power plants require billions in compliance costs; that is why there are so few of them. If AI licensing requires the kind of safety infrastructure only Anthropic has built — because only Anthropic has been positioning itself as the safety company — then the license is not a public safety measure. It is a barrier to entry wearing a public‑safety label.
Doctorow’s term for this — borrowed from Jay Freeman — is the “felony contempt of business model”: the legal architecture that criminalises competition with an incumbent’s way of making money. In the AI case, the architecture is not criminal law but regulatory capture of a familiar kind: a safety‑licensing structure that only the company loudly warning about safety can satisfy. The irony, if it is irony, is that the company calling for regulation is the company best positioned to benefit from it.
None of this requires Anthropic’s leadership to be cynical. It is entirely possible that Dario Amodei genuinely believes his models pose an existential risk and is genuinely trying to prevent catastrophe. It is also entirely possible that both things are true simultaneously: the models are dangerous, and the danger narrative is worth a trillion dollars. Galbraith’s concept of the bezzle — the interval when a confidence trickster knows he has the money but the victim does not yet understand the loss — applies here in a modified form. The bezzle is not fraud. It is the gap between what is claimed and what can be verified, during which the claimed value circulates as if it were documented value. Anthropic claims its models can conduct end‑to‑end cyberattacks. Anthropic claims its models show early signs of recursive self‑improvement. No one outside Anthropic can check either claim. The IPO will price them as engineering facts.
Russell writes that “unrestrained development of unsafe systems leads to intolerable risks.” He may be right. But the alternative he proposes — a licensing regime built on the self‑assessments of the companies seeking to be licensed — has a structural defect he does not name. The regulated and the regulators would work from the same unverifiable capability claims, produced by the same companies, about systems whose internals are proprietary. It is as if the nuclear licensing regime depended on the reactor manufacturer’s own assessment of the reactor’s safety, because only the manufacturer understood the physics. If Anthropic’s models are as dangerous as Anthropic says, then the public has an interest in independent verification of that claim — not the company’s word, taken on faith and priced into an S‑1 filing. Safety evaluations of frontier models should be conducted by parties with no financial interest in the outcome, and the methodology should be public. Anything less is not regulation. It is marketing.
Russell’s proposed licensing regime, built on the same unauditable claims that will underwrite the IPO, turns public deliberation into a pre‑offering roadshow. The consultation is not a public comment period. It is an S‑1 filing. The deadline is the IPO date.