Anthropic CEO Dario Amodei has published a long piece arguing the time has come to slow AI development. At its centre sits a three-step plan.
He calls it pacing the frontier: lowering the tempo of training and development to give companies time to build safeguards and regulators time to evaluate models.
The three steps
The staging is deliberate. Each step demands broader participation than the one before it.
| Step | Scope |
|---|---|
| Embedded auditors | Permanent, independent auditors inside the companies |
| Shared standards | Common safety criteria among democratic nations |
| Global treaty | An international agreement that includes China |
Anthropic is taking the first step unilaterally, now. The company says it will give third-party evaluators such as METR wide-ranging access to its models, so that adherence to its safety practices and commitments can be verified from outside.
What an embedded auditor means
Anthropic's commitment goes further than granting access. The company says it will accept permanent auditors with access to internal systems and the right to publish their findings.
Amodei wants governments to mandate this across the industry. That is the distinction: a company's voluntary transparency and a legally defined audit obligation are not the same thing, and without the second the first turns into a competitive disadvantage.
The embedded auditor idea is also a test. An auditor being able to publish findings means the company accepts results that may go against it; the seriousness of the commitment will be measured by the first unfavourable report.
How the treaty is framed
The third step includes a four-tier structure, topped by a speed limit on recursive self-improvement. Amodei compares the approach to Cold War arms reduction agreements.
His case for urgency is concrete: he argues systems capable of autonomous cyberattacks could threaten the entire internet within six to twelve months, citing documented incidents at OpenAI and Anthropic.
An agreement including China is the hardest part of the plan. In arms treaties verification rested on physical inspection; what the equivalent would be in software finds no clear answer in the text.
What the bought time is for
Amodei stresses that slowing down is not the goal in itself. The time gained, he says, should go to safety research, interpretability work and stricter testing.
His analogy is aviation, where safety standards took decades to settle. That is also the weak point of the piece: in aviation the standards were shaped by crashes, whereas here an agreement is being sought in advance, against a risk that has not yet materialised.
Why the timing draws attention
The piece arrived just before the company's IPO, expected in November and possibly the largest in history. In the same week it was reported that Nvidia is in talks to put up to 10 billion dollars into that offering.
That does not make the call insincere, but it does change how it reads. A company asking for a slowdown while preparing a record capital round will find that fact shaping how the proposal is received.