Key Highlights

  • Anthropic has selected Accenture as its first embedded third-party evaluator, marking the initial concrete step in implementing a proposed framework to slow the pace of AI development.
  • Anthropic and Accenture have agreed to invest at least USD 1 billion in building capacity in this area over the next five years.
  • Embedded evaluators from Accenture's AI Business will be given employee-level access to test safeguards, red-team models and assess alignment with human values.
  • Anthropic said it is also in discussions with a research nonprofit and other potential third-party evaluators.

A First Step Toward a Broader Framework

Anthropic has selected Accenture as its first embedded evaluator, marking the company's initial concrete step toward implementing a proposal from its chief executive to slow the pace of AI capability development. Anthropic and Accenture have agreed to invest at least USD 1 billion in building evaluator capacity over the next five years, with Anthropic funding Accenture's work directly given what the company described as the importance and urgency of the effort.

Anthropic said that longer term, it believes funding for such evaluation work should come from pooled or government sources, as it previously called for in a policy framework published earlier this year, and that in the absence of such funding mechanisms, it plans to work with different evaluators under varying funding arrangements.

The Broader Slowdown Proposal

The announcement follows a proposal published by Anthropic's chief executive outlining a three-step plan intended to temper the pace at which AI companies improve their most advanced models. Anthropic and a chief industry rival have faced intense scrutiny in recent weeks following warnings from researchers about the potential for AI to cause catastrophic harm, with the proposal drawing mixed reactions from industry leaders, some supportive and others dismissive of the need for new safeguards.

The first step of the proposal involves granting third-party evaluators employee-level access to verify safety practices and report incidents, a step Anthropic said it has committed to unilaterally while encouraging other AI companies to adopt similar practices.

Scope of the Partnership

Anthropic said it has initially agreed to embed employees from Accenture's specialist AI business into the company to test safeguards, red-team models, and assess whether systems behave in line with human values. The company emphasized the Partnership is not exclusive and said it is in discussions with a research nonprofit focused on AI safety, along with other potential third parties, while stressing that working with embedded evaluators does not reduce Anthropic's own accountability for the safety of its models.

Outlook

Anthropic said its approach to embedded evaluation is expected to evolve as the field matures, with the company indicating it will share further details as its work with Accenture and any additional evaluators progresses.