TL;DR: Anthropic lets six METR researchers full‑access to its internal tools, framing it as an “independent audit” akin to bank examiners, but the move may be more PR than substantive oversight.
The Details
Anthropic announced that six researchers from the nonprofit METR will receive employee‑level laptops, office space, and access to internal risk‑assessment tools. The company says the arrangement is meant to close the gap between external evaluators’ understanding and the company’s internal view, and it likens the setup to banking regulators placing examiners inside banks. METR will publish its own assessments and Anthropic cannot veto publication, only request limited redactions for customer data, attorney‑client privilege, or model architecture details. The article notes a wave of AI incidents since August, fueling calls for independent checks, but provides little evidence that this access will materially change Anthropic’s safety practices.
Watch For
- The first METR assessment report and whether any redactions spark disputes.
- Any regulatory bodies (e.g., FTC, EU AI regulators) referencing this arrangement in future guidance.
- Reactions from other AI auditors or competitors that could signal broader industry shifts.