The Information Machine
Following·Day 2·first covered 11 Sep 2026·6 sources

Anthropic and OpenAI Both Commit to Embedded Third-Party AI Evaluators

The gist

Two leading frontier AI labs have made explicit, operational commitments to third-party access that goes beyond prior voluntary safety pledges. The antitrust waiver request and multi-stage coordination plan, if pursued, would require government action to take effect.

The full picture

Anthropic and OpenAI have each committed to granting independent evaluators permanent, employee-level access to verify safety practices, report incidents, and assess model alignment during training runs. Dario Amodei published an essay titled 'We Must Pace the Frontier' proposing a three-stage plan: first, embedded evaluators at each frontier lab; second, competing labs coordinating on shared safety standards with a narrow antitrust waiver from the US government; third, democratic governments coordinating with authoritarian ones on AI pacing. Anthropic is unilaterally committing to the first stage now, with Amodei citing organizations like METR as examples of evaluators and the banking industry as a precedent for the access model. Practically, Anthropic plans to give reviewers desks, badges, laptops, and permissions comparable to internal risk-assessment teams, with carve-outs where law, contracts, or customer privacy require them. Sam Altman endorsed Amodei's essay and stated OpenAI will match the embedded evaluator commitment, providing evaluators with permanent, employee-level access. In a separate Fortune Magazine interview, Altman responded to a question about whether lab leaders could coordinate on AI development by saying 'I think that will happen.' Amodei also warns that within 6 to 12 months an AI-powered botnet could cause hundreds of billions of dollars in damage, and argues any US slowdown must not exceed the lead American companies hold over Chinese AI projects. He proposes chip export restrictions and protections against model theft as mechanisms to maintain that lead while allowing domestic pacing. John Schulman has separately called on OpenAI and Anthropic to stop feuding and develop a joint pacing proposal, noting the Noerr-Pennington doctrine means antitrust law does not bar jointly developing such a proposal. The Senate committee with jurisdiction over commerce and technology policy has no AI hearings scheduled for the next four months.

How it developed
12 September 2026

Altman said in Fortune Magazine interview that cross-lab coordination among major AI leaders 'will happen'

11 September 2026

John Schulman called on OpenAI and Anthropic to jointly develop a pacing proposal, citing Noerr-Pennington doctrine; Senate committee noted to have no AI hearings for four months

Sources
1 more source
The daily email

Want this in your inbox?

I send one email each morning with the stories that moved. If you would rather just read here, that works too.

Subscribe free