Anthropic lets outside evaluators examine its AI systems
What's the story
Anthropic, the company behind the Claude AI models, has given non-profit organization Model Evaluation and Threat Research (METR) employee-level access to its systems. The move is part of a broader effort by AI firms to involve external evaluators in their work. This way, they can safely advance technology without hindering its development.
Access details
METR's access to Anthropic's systems
Anthropic has given six employees from METR access to company hardware, office space, and internal systems. This is similar to what its risk assessment teams have.
The company said that while it can't veto METR's findings on risks or practices, it can ask for narrowly defined redactions. These could include customer data, legally privileged information, and model architecture details.
Transparency commitment
METR's report will detail the evaluation process
In line with its transparency policies, METR has said that its report will clearly outline the scope, access, and other limitations of its work.
The non-profit had suggested that developers allow private staff interviews, timely publication of assessments, and clarity on access limits and redactions.
This is part of a broader discussion about the balance between AI safety and development speed.
Access tiers
New 3-tiered access system for evaluators
Anthropic is also working on a three-tiered access system for evaluators.
The first tier will conduct narrowly scoped investigations with approval needed to publish.
The second tier will have more information and can publish without approval, but subject to redactions that evaluators can publicly challenge.
The third tier, which METR is using, gives employee-like access with similar publication rights as the second tier.
Ongoing evaluations
SecureBio is in first tier of this access system
SecureBio, a biosecurity non-profit, is in the first tier of this access system. It is currently reviewing the chemical and biological weapons chapter of Anthropic's August risk report.
The company has said it will consider expanding SecureBio's access and publication rights after this review.
This shows Anthropic's commitment to transparency and collaboration in AI safety efforts.