Anthropic has implemented a concrete measure to begin applying the proposal of its CEO, Dario Amodei, to slow down the development pace of more advanced artificial intelligence (AI) systems by incorporating external evaluators into the company.
Accenture was chosen as the first embedded external evaluator. This collaboration will allow specialists from Faculty, Accenture's AI division, to work internally at Anthropic to examine the company's safety mechanisms. Activities include safeguard testing, red teaming exercises, and analyses aimed at confirming that the AI models adhere to human values.
Anthropic and Accenture have established a minimum investment of US$ 1 billion (approximately R$ 5.3 billion) to develop this capability over the next five years. However, Anthropic stated that, given the urgency and relevance of the work, it will fund Accenture's operations itself.
In the long term, the company expects such activities to be funded by governmental or shared resources, as outlined in the Advanced AI Framework presented by Anthropic in June. Since these structures have not yet been established, the company plans to collaborate with various evaluators and funding models.
Amodei's suggestion proposed that independent evaluators have access to the employee level within AI companies. The purpose of this is to empower these specialists to directly verify safety practices and report any incidents.
When presenting his idea, Amodei assured that Anthropic would undertake this initial stage 'unilaterally' and encouraged other industry corporations to adopt a similar stance.
Other partnerships and responsibility
The collaboration with Accenture will not be exclusive; Anthropic informed that it is also in contact with the non-profit research organization METR and other potential entities to participate in this process.
Despite this, the company emphasized that the responsibility for the safety of its models remains entirely under its purview, and the presence of external evaluators does not diminish this duty.
Anthropic commented: "We are sharing these initial efforts now so that people and other AI developers can see our process." The company added that it intends to refine this methodology as the sector matures and will communicate new information as soon as the work begins and more evaluators are integrated.
With Accenture's participation, the company took a fundamental practical step to convert one of Amodei's central plan ideas into a functional structure: allowing an external entity direct access to audit its safety systems and detect potential failures before they escalate.
