Anthropic is giving six researchers from the nonprofit METR employee-level access to independently assess its AI models, safety practices and risk management. The arrangement allows METR to publish its own assessments of risks, incidents and company practices without Anthropic being able to block publication.
Internal tools and information from Anthropic staff
The METR team has company laptops supplied by Anthropic and office space. Its access to internal tools is roughly equivalent to that available to Anthropic’s own risk assessment teams.
The arrangement also covers sharing internal information. Anthropic says it encourages employees to share information with METR researchers as they would with a colleague studying the risks of AI behavior diverging from human intentions.
Publication without a veto, subject to defined redactions
METR will publish its own assessments of Anthropic’s risks, incidents and practices. The company has waived its right to stop publication but retains the ability to require redactions of information within narrowly defined categories.
Anthropic says those categories include customer data, material protected by attorney-client privilege and information about its model architectures. METR may publicly describe the requested redactions and discuss whether it considers them justified, provided it does not reveal the redacted information.
METR is the only evaluator in the highest tier
Anthropic is building a three-tier external evaluation system, and METR is currently the only organization in its highest tier, which adds employee-level access. The middle tier gives evaluators more information and allows them to publish without prior company approval.
The lowest tier includes SecureBio, a nonprofit focused on biosecurity. Evaluators in this group conduct narrowly scoped assessments and need Anthropic’s approval to publish them.
Putting Amodei’s external evaluator proposal into practice
The new arrangement implements a proposal Anthropic CEO Dario Amodei outlined in a September essay. He proposed embedding external evaluators directly within AI labs as a first step toward slowing the technology’s development.
