Anthropic has officially announced that it has begun collaborating with the nonprofit organization METR to provide employee-level access to its artificial intelligence models for more robust independent evaluations. This initiative is part of the company's commitment last month to enhance independent third-party assessments of its models and safety practices.
Reasons for Collaborating with Independent Evaluators
As the capabilities of models continue to grow, the demands for safety measures and risk assessments have also increased. Anthropic stated that they aim to be accountable for their decisions, but cannot publicly share all information related to safety work, as some information could be exploited by malicious actors or is commercially sensitive data.
“We want to narrow the gap between evaluators' understanding of cutting-edge AI systems and the company's own understanding.”
Anthropic
Role of METR and Evaluation Methods
METR will be responsible for assessing Anthropic's risk reports and publicly disclosing its risk assessment results. Anthropic has already collaborated with METR on multiple reviews of risk reports and plans to further strengthen partnerships with other evaluation agencies, including SecureBio and Accenture.
“Our goal is to ensure that our partners can protect confidential information and possess the resources and expertise necessary to evaluate our models.”
Anthropic
Access Levels for Evaluators and Their Impact
Anthropic plans to establish three access levels for evaluators, gradually increasing the information access rights for evaluators as the collaboration deepens. Level one evaluators will conduct narrow risk investigations, while level two evaluators will be able to publish results without company approval. The collaboration with METR falls under level three, granting access rights similar to those of employees.
“We believe that every cutting-edge AI company should adopt similar embedded evaluation measures.”
Anthropic
This series of measures marks an important step for Anthropic in AI safety regulation, aimed at enhancing transparency and accountability for its models.
Source: Anthropic Official Announcement