Anthropic invites the examiners inside

Calls to slow down artificial intelligence are cheap and frequent. What makes this one worth reading is that it comes with a mechanism, and the mechanism is borrowed from banking supervision rather than from the language of safety pledges.

Dario Amodei published an essay on Saturday titled We Must Pace the Frontier, setting out a three-part plan and committing Anthropic unilaterally to the first part. The Guardian reported the proposal: third-party evaluators would get permanent, employee-level access to the company's systems to verify adherence to its safety measures, report incidents and assess alignment during training.

The detail is what distinguishes it from an undertaking. TechCrunch reported that evaluators from outside organisations such as METR would receive company badges, desks and laptops, with access mostly comparable to what internal risk assessment teams hold, excepting what law or contract forbids.

Amodei's own comparison is to the regulators who sit inside banks. That is a precise analogy and a demanding one. A supervisor with a desk sees what a quarterly attestation cannot, which is the ordinary conduct of the firm on days when nothing is being reported.

He frames the commitment as unilateral and asks governments to require other frontier companies to match it. The second half of that sentence is the operative one: a voluntary standard adopted by one firm is a competitive disadvantage until it is mandatory.

The response came quickly. Sam Altman said he agreed that the industry needs to pace the frontier, called independent evaluators with employee-like access a good idea, and said OpenAI would do the same, describing it as a primary topic of discussion at the company in recent weeks. Elon Musk posted that Dario is right.

The second part of the plan is harder and less discussed. Amodei called for leading companies in democratic countries to coordinate on common safety standards and on limits to the rate of unchecked progress, which is an agreement between competitors about how fast to compete.

He is explicit about the legal obstacle. Firms have reportedly worried that a coordinated pause invites antitrust scrutiny, and he asks the American government to mediate or at least enable the discussions, issuing a narrow waiver for certain kinds of safety conversations.

That request is the most consequential paragraph in the essay, and the one least likely to survive contact with Washington. A waiver permitting rivals to agree on output restraint is a significant thing to hand an industry, however the restraint is described.

On the argument that slowing down cedes the field to China, Amodei's answer is that export controls on advanced chips and semiconductor manufacturing equipment, together with a crackdown on model distillation, change the arithmetic. It is a case for pacing the frontier while restricting who else can reach it.

The timing invites scepticism and deserves it. The essay arrived days after a former Anthropic researcher, Jacob Coxon, resigned saying neither Anthropic nor OpenAI was acting responsibly, a day after the company published its own account of misuse it had blocked, and weeks before an intended listing. All three can be true at once, and the proposal still has to be judged on whether anyone signs a badge out.