METR evaluators gain access to Anthropic’s offices and internal systems
Anthropic has embedded six independent evaluators from nonprofit METR in its offices to assess the risks of its AI models. The move puts into practice a proposal for outside auditors that Anthropic chief executive Dario Amodei and rival technology leaders supported, as concerns grow about the effects of increasingly capable AI.
The METR team has company hardware, office space and access to internal tools and systems comparable to those used by Anthropic’s risk teams. METR can publish its findings without Anthropic’s approval, though the company may require redactions for customer data, privileged material or model architecture details. Anthropic says it is not paying METR; OpenAI has not yet announced a similar arrangement.
- Six METR evaluators are embedded at Anthropic.
- They can publish findings without company approval.
- OpenAI has not announced a similar plan.
New here? Start with this
Anthropic is a company that develops artificial intelligence models, including Claude, a system that can respond to questions and carry out tasks. As these models become more capable, researchers and policymakers are examining whether they could cause harm, for example by helping people carry out cyber attacks or by behaving in ways their developers did not expect.
METR is a nonprofit research group that evaluates advanced AI systems. Independent evaluators can examine how a company tests and manages its models, giving outsiders a way to assess claims about safety. The arrangement also raises questions about how much access auditors need and what information can be made public while protecting sensitive data.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Giving METR access to Anthropic’s offices and internal systems could make its assessment more informed than an external review based only on material the company chooses to share. METR’s ability to publish findings without company approval may strengthen public accountability, while limited redactions can protect customer information and genuinely sensitive details. The arrangement also puts into practice a proposal for independent scrutiny as AI capabilities and potential risks grow.
The case against
Embedding evaluators in a company’s offices and giving them tools comparable to those used by its own risk teams may make it harder to maintain the distance expected of an independent auditor. Redactions for privileged material or model architecture could also limit what the public learns, even if METR retains publication rights. Without a similar arrangement announced by other leading developers, one company’s review may offer only a partial view of wider AI risks.
Read the full article at the source →
Originally published by Ars Technica as “Anthropic gives outside AI watchdog “employee-like” access”.