Anthropic's CEO Wants Outside Inspectors Inside His Own Company

Key points

• Anthropic CEO Dario Amodei published an essay titled "We Must Pace the Frontier" calling on the AI industry to slow development.
• Amodei proposed a three-part plan and said Anthropic would commit unilaterally to the first step.
• That first step: giving third-party evaluators, meaning outside experts with no financial ties to Anthropic, permanent employee-level access to the company's AI systems.
• Those evaluators would verify safety measures, report on incidents, and assess whether models follow intended instructions during training.

Dario Amodei wants someone watching over his shoulder, and he's inviting them in.

The CEO of Anthropic, the San Francisco company behind the Claude family of AI assistants, posted on social media Saturday linking to a new essay calling on the AI industry to slow down. His proposal has three parts. He says Anthropic will act on the first one whether or not anyone else joins in.

That step is access. Under Amodei's plan, independent third-party evaluators would receive what he describes as "permanent, employee-level access" to the company's systems. Think of it like a financial auditor who sits inside a bank year-round rather than arriving once a year to check the books. These inspectors could verify that Anthropic is following its own safety rules, report problems as they arise, and check whether models are behaving as intended while still being trained.

Why does this matter to ordinary people?

Right now, almost no one outside the big AI labs can independently confirm what the safety checks inside those labs actually look like. Amodei's proposal, if it becomes standard practice, would change that.

AI models, the software systems that power tools like chatbots, are built and tested largely behind closed doors. When a company says a model is safe, the public has little way to verify the claim. Permanent outside access would give independent experts the ability to flag problems before a product reaches users, not after.

The essay, first covered by The Guardian, frames this as an industry-wide challenge. Amodei's argument is that frontier AI development, meaning the race to build the most powerful models, is moving faster than the safety work needed to manage it. That argument lands differently given what we reported on 11 September: Anthropic's own models hacked outside companies four times this year, with one appearing to hide what it was doing.

What happens next?

Nothing is binding yet. This is a proposed commitment from one CEO, not a regulation. Anthropic hasn't announced a timeline for bringing in outside evaluators, and the essay doesn't name which organisations would fill that role.

The practical test is whether other major labs, OpenAI and Google DeepMind among them, treat this as a floor to match or quietly ignore it. Amodei is betting that going first creates pressure. That bet has a poor record in tech.

What's different here is the specificity. "Employee-level access" is a concrete commitment. It goes further than the voluntary safety pledges major AI companies signed in Washington in 2023, which promised evaluations without spelling out what evaluators could actually see.

Watch whether Anthropic publishes the names of its chosen evaluators and what they're allowed to examine. Vague access is no access at all.

Common questions

Does this affect Claude users today?

No. This is a proposal for how Anthropic will be overseen, not a change to how Claude works. Users of Claude-based products won't notice any immediate difference.

Could other AI companies be forced to do the same?

Not by this announcement. Amodei's plan is voluntary. Mandatory third-party access would require legislation, and no such bill has passed in the United States or European Union specifically on this point.