Anthropic gives third-party auditors ’employee-like access’ in AI industry first

Anthropic disclosed Thursday that it will grant "employee-like" access to six evaluators from METR, a third-party nonprofit specializing in AI model testing. These evaluators will be provided with Anthropic hardware, office space, and access to internal tools and systems comparable to those used by Anthropic's own risk assessment teams.

The evaluators are permitted to share information with them, similar to how they would share information with a colleague on alignment risk. Anthropic has specified that certain categories, such as customer data, attorney-client privileged information, or insights into model architectures, should not be shared. METR will publish its findings, which Anthropic is not permitted to withhold, although the company may request the redaction of sensitive information as necessary.

Anthropic CEO Dario Amodei had previously proposed embedding evaluators from nonprofits like METR in a September blog post, proposing them as one of three measures to address "frontier" AI development. Anthropic expressed a commitment to accountability, acknowledging that "not all relevant information to our safety work can be publicly shared" due to potential risks associated with malicious actors or commercial sensitivity, such as details about compute or training information.

By allowing METR to gain deep access to Anthropic's models and company information, Anthropic aims to bridge the gap between the evaluators' understanding of the systems and the company's own knowledge of these systems. As US AI regulation has been relatively lackluster to date, some researchers and advocates have advocated for third-party reviews of models, either pre-release or throughout the development process.

The FRONTIER Act, a bipartisan bill introduced in July, would mandate AI companies to allow independent verification organizations, or IVOs, to assess the safety of their development process. Initial support for the bill was expressed by OpenAI, marking the first instance of such backing for proposed federal AI regulation. However, the lack of binding nature of these voluntary pledges leaves the AI industry largely to self-regulate, as evidenced by recent developments including a subpoena from the FTC into OpenAI and Anthropic for concerns over consumer harm.