Anthropic gives METR researchers employee-level access to assess its AI models

[EdgeLord]

Anthropic is giving six researchers from METR employee level access so they can independently assess its AI models and safety practices. TechCrunch's Aditya Mehta reports that METR is a nonprofit that evaluates frontier AI models. It is one of the most open arrangements between an AI lab and an outside evaluator so far

The researchers will get company laptops, office space and internal tools roughly equivalent to those used by Anthropic's own risk teams. Anthropic staff are encouraged to share information with them as they would with a colleague. METR will publish its own independent assessments of risks, incidents and practices

Crucially, Anthropic has given up any veto over what METR publishes. It can require redactions only in three narrow areas: customer data, legally privileged material and details of model architecture. METR is allowed to say publicly whether it thinks those redactions were justified, which adds another layer of accountability

METR is the only evaluator in the top tier of a new three tier system. Lower tiers include organisations such as SecureBio, which need Anthropic's approval before publishing. The move follows weekly disclosures of AI agent incidents since August and an essay by chief executive Dario Amodei in September calling for evaluators embedded inside AI labs

Independent oversight from people with real access is something many safety researchers have asked for. Giving up the veto on publication is a meaningful commitment. If it works well, other labs may feel pressure to offer similar arrangements

Is embedded independent evaluation the right model for AI oversight? And should other labs be expected to follow?

Anthropic is letting METR researchers look under the hood with employee-level access | TechCrunchtechcrunch.com