Anthropic and Accenture Partner on Embedded Evaluation for Frontier AI
Anthropic has partnered with Accenture to implement "embedded evaluation," a process where independent evaluators work inside the AI company to verify safety commitments and assess model development. This partnership aims to increase the verifiability of AI safety by providing external experts with employee-level access to training processes and internal decision-making.
The Embedded Evaluation Model
Embedded evaluation differs from traditional external evaluation by granting independent evaluators access comparable to that of an employee. This internal vantage point allows evaluators to:
Monitor Model Development: Watch models take shape during the training phase.
Audit Internal Governance: Follow the decisions that govern how models are built and deployed.
Direct Employee Engagement: Speak directly to employees to identify blind spots and assess operational safety.
Public Reporting: Report incidents and provide the public with a more informed account of the risks and benefits associated with the models.
Anthropic states that while these evaluators provide verification, the ultimate responsibility for model safety remains with the AI developer.
Partnership Scope and Investment
The partnership is led by Faculty, Accenture's specialist AI business. The scope of work includes evaluating and red-teaming models, conducting alignment assessments, and testing model safeguards. Accenture's experience in deploying AI across various industries and governments is intended to inform the safety approach used during these evaluations.
Both Anthropic and Accenture expect to invest at least $1 billion each over the next five years to build capacity in this area.
Funding and Standardization Challenges
Because there are currently no established standards for information access or reporting for embedded evaluators, and no settled system for funding, the current arrangement is as follows:
Direct Funding: Anthropic will fund Accenture's work directly due to the importance and urgency of the initiative.
Alternative Funding Pilots: Anthropic is in dialogue with METR and other nonprofit evaluators to pilot embedded evaluation using their own independent funding.
Long-term Goal: Anthropic advocates for funding to eventually come from pooled or government sources, as proposed in their Advanced AI Framework published in June.
Ecosystem and Non-Exclusivity
This partnership is non-exclusive. Anthropic intends to work with multiple evaluation organizations, with more to be announced in the coming weeks. Similarly, Accenture will work with other AI developers in similar capacities. The goal is to establish an ecosystem of evaluators operating under shared standards as the field of frontier AI matures.