TL;DR
- Dario Amodei announced on 12 September that Anthropic will embed independent evaluators with employee-like access, the first step of his plan to slow frontier AI development while safety work catches up.
- Sam Altman said OpenAI will do the same. He did not name an evaluator, publish access or reporting terms, or give a start date.
- Elon Musk endorsed Amodei’s proposal but made no commitment for xAI. The three posts were separate public statements, with no published joint plan or timetable.
In plain English
Pacing the frontier: slowing improvements in advanced AI capabilities so safety work has time to keep up, without stopping model development.
Embedded evaluator: an independent reviewer working inside an AI company with access comparable to its internal risk-assessment staff.
Employee-like access: ongoing access to the people, tools and permissions that a company’s own staff doing comparable risk assessments have.
OpenAI CEO Sam Altman said on 12 September that the company will give independent evaluators employee-like access. His announcement adopted the first step of a three-part plan published earlier that day by Anthropic CEO Dario Amodei to slow frontier AI development while safety work catches up.
Amodei called the plan “pacing the frontier.” Anthropic committed to embedding third-party evaluators inside the company. They would receive desks, badges, company laptops and access mostly comparable to internal risk-assessment teams, with exceptions where the law, contracts or customer privacy require them. Their contracts would allow them to publish key findings without editorial control from Anthropic. Anthropic would keep a narrow right to redact security-sensitive, legally privileged, commercially sensitive or third-party confidential information.
The second step calls for frontier AI companies in democratic countries to agree common safety standards and limits on the pace of progress. The third calls for democratic governments to try to coordinate with authoritarian governments. Amodei wrote that the steps do not need to be taken strictly in order. Anthropic can implement embedded evaluation on its own. The other two steps require other companies and governments to participate.
Altman wrote: “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon.” His post did not name an evaluator, describe its access or reporting rights, or give a start date.
Elon Musk separately endorsed Amodei’s proposal with the words “Dario is right.” His post made no commitment for xAI. The three posts show public support for Amodei’s proposal; there is no published joint plan or timetable between the companies.
My view: OpenAI’s public commitment is useful progress. Its implementation becomes measurable when OpenAI names the evaluator and publishes the access, independence, reporting rights and timing. Those details will show whether “we will do the same” matches Anthropic’s commitment in practice.