OpenAI Backs Anthropic’s Call for OpenAI Embedded Independent Evaluators

OpenAI CEO Sam Altman has committed the company to the Dario Amodei embedded evaluators proposal—an OpenAI Anthropic paced frontier safety commitment—according to the Associated Press. The alignment puts two leading frontier-model developers behind a more intrusive form of outside safety review than one-off audits. Associated Press

Amodei’s proposal is not simply that companies publish more safety claims or commission occasional external reports. In his essay, he argues that frontier AI developers should provide third-party evaluators—including organizations such as METR—with continuing access comparable to that available to employees. The purpose, Amodei writes, is to let evaluators verify safety practices, assess training processes and report incidents. Dario Amodei, We Must Pace the Frontier

The Associated Press says Altman quickly agreed that OpenAI would adopt the same kind of independent embedded-evaluator arrangement, while also supporting Amodei’s argument that AI development should slow enough for safety measures to catch up. Associated Press

That is a meaningful public commitment, but it should not be mistaken for a completed governance system. Neither the AP account nor Amodei’s essay, as described here, establishes what access OpenAI will provide, which evaluators it will select, whether their findings will be public, or what authority they will have when they identify a serious problem. The sources also do not establish a shared binding standard across the industry, a specific deployment threshold, or a timetable for slowing model development.

Why embedded access matters

The distinction between an embedded evaluator and a conventional outside assessment is material. Amodei’s model asks for ongoing access rather than a narrowly scoped engagement designed around a single release or benchmark. According to Amodei, evaluators should be able to inspect safety practices and training processes as they unfold, as well as report incidents. Dario Amodei, We Must Pace the Frontier

That could reduce one recurring weakness of voluntary AI-safety commitments: companies often control the timing, scope and disclosure of evidence offered to outsiders. Continuing access could make it harder to treat safety evaluation as a launch-stage formality. But the public information cited does not say whether evaluators would be able to inspect all relevant systems, publish adverse conclusions without company approval, or trigger a pause in training or deployment. Those details determine whether “employee-like access” becomes meaningful oversight or remains a limited advisory role.

METR is a concrete reference point in the announcement. Amodei names it as an example of the kind of third-party evaluator he believes frontier AI companies should embed. Dario Amodei, We Must Pace the Frontier OpenAI has also already cited METR in a separate context: its account of the Hugging Face incident says that METR and Redwood Research each conducted and published an independent investigation into alignment issues connected to that incident. OpenAI, The Hugging Face incident and the road ahead

That prior work shows OpenAI has used outside investigators and allowed their work to be published in at least one instance. It does not, by itself, demonstrate that OpenAI has already implemented the permanent access model described by Amodei. Nor does it show that the two companies have agreed on common evaluation methods or reporting rules.

A useful commitment, awaiting testable terms

The strongest conclusion supported by the public record is narrow but important: Anthropic has called for persistent third-party scrutiny of frontier AI development, and OpenAI’s chief executive has publicly backed the same arrangement. Dario Amodei, We Must Pace the Frontier Associated Press

For developers and researchers, the next evidence to watch is operational rather than rhetorical: named evaluators, defined access rights, published methodologies, incident-reporting rules and disclosures that can be independently checked. Until those terms are public, claims about how much this will constrain frontier development remain unproven.

The proposal is therefore neither a new regulatory regime nor a demonstrated slowdown in AI capability work. It is a public move toward a more accountable evaluation model—one whose credibility will depend on whether independent reviewers can see enough, say enough and publish enough to make their independence consequential.