Anthropic CEO Calls for Slowing AI, Starting With Deeper Outside Oversight
Anthropic chief executive Dario Amodei called on September 12, 2026, for the artificial intelligence industry to slow down. He announced an essay titled We Must Pace the Frontier in a social media post, according to The Guardian.
The announcement was published as a post on the @DarioAmodei account on X at 2:01 PM, linking to darioamodei.com/post/we-must-pace-the-frontier. That post had 4.8 million views, per the X post.
Amodei offered a three-part plan for slowing AI progress. Anthropic said it would commit on its own to the first step of that plan. That step centers on third-party oversight. It would give external evaluators permanent, employee-level access to Anthropic systems to verify safety measures, report incidents, and assess alignment during training. Alignment here means whether a model's behavior stays consistent with human intent and safety rules.
Amodei wrote that the pace at which AI model capabilities are improved must be slowed so risk prevention has time to keep up. He has said artificial intelligence will surpass human cognitive abilities across most tasks within a "small number of years." Former Anthropic researcher Jacob Coxon warned that AI could precipitate human extinction by 2030.
There is deep distrust between Amodei and OpenAI chief executive Sam Altman, as reported by Reuters. An official identified as Lutnick sent a letter to Amodei ordering the company to suspend export of its AI models, according to Reuters.
The broader context here is commercial rivalry and state power already intersecting with frontier development. Frontier means the most advanced systems. It is a collective-action dilemma layered over a verification dilemma. A unilateral promise can set a procedural example without removing the incentive for others to move faster. Rival labs face shareholder pressure, national security expectations, and fear of falling behind. Voluntary restraint is fragile under those conditions. Amodei's answer appears to be transparency by design, making safety claims checkable rather than taken on trust. Permanent evaluator access would narrow the information gap between developer, auditor, and eventually regulator, like giving building inspectors a permanent site badge instead of a one-time tour. It also tests whether continuous assessment during training can catch alignment failures earlier than testing only before release.
Looking at what this means for governance, two vectors matter. The first is whether evaluator access becomes a shared norm. If other frontier labs adopt equivalent access, it could form the technical basis for mutual assurance or treaty-linked inspection ideas. If they do not, Anthropic carries the cost and friction alone. Trust is scarce. The reported distrust between Altman and Amodei shows how personal and institutional rifts complicate any pact on pacing, even when the stated risks overlap.
A second question for governance concerns the state. Export suspension orders are blunt instruments. They assert control over where models spread. Evaluator regimes assert a different control, over knowledge about how systems are built. The two logics can reinforce each other or collide. Governments seeking leverage may welcome deeper audit trails. Developers seeking deployment latitude may find that openness invites more intervention. For experts watching international institutions, the question is which model scales across borders, where legal authority, technical capacity, and strategic interest diverge sharply.
