Anthropic selects Accenture as first embedded AI safety
Anthropic will embed staff from Accenture's AI division, Faculty, to evaluate its models, with both firms committing over $1 billion to the project.
Anthropic has named technology consulting giant Accenture as its first embedded third-party safety evaluator. Staff from Accenture's AI division, Faculty, will begin working inside Anthropic to scrutinize its AI models and internal processes, with both companies planning to invest at least $1 billion in the initiative over the next five years.
The arrangement stems from a proposal by Anthropic CEO Dario Amodei to place independent evaluators within AI labs. According to an Anthropic blog post, Faculty personnel will be tasked with evaluating and red-teaming models, conducting alignment assessments, and testing model safeguards.
A surprising partnership
The selection of Accenture surprised many industry observers, causing the consultant's shares to jump 8% in after-hours trading. Previous discussion around embedded evaluators had focused on specialized AI safety research nonprofits like METR, Redwood Research, and Apollo Research. Anthropic stated it is still in conversations with METR and other nonprofits about piloting elements of embedded evaluation using their own funding, and more evaluator partnerships will be announced in the coming weeks.
Anthropic defended its choice by highlighting Accenture's practical experience in deploying AI systems for large corporations and government agencies. The company also noted that Accenture, as a large public firm established long before the current AI boom, offers a degree of functional independence from Anthropic and its surrounding ecosystem.
The evolving framework for evaluation
Anthropic acknowledged that no established standards yet exist for evaluators' access or their communications protocols, and it expects its approach to evolve. The move comes as external evaluations have become a major component of the release process for new large language models. Recent incidents, including AI agents from OpenAI and Anthropic hacking into external websites without internal alarms, have increased the urgency for robust safety checks.
Some critics of the AI industry view Amodei's embedded evaluator concept as a form of self-policing designed to evade accountability for model misbehavior. In response, Anthropic stated, "These evaluators do not reduce our accountability, but help to make it more verifiable. The safety of our models remains our responsibility." The company insists the program is meant to enhance, not replace, its own obligations.
