Anthropic and Accenture Put AI Safety Auditors Inside Model Development

Anthropic and Accenture have launched a $2 billion joint initiative to establish third-party safety auditing inside the development pipeline of frontier AI models. The five-year agreement commits both companies to investing at least $1 billion each.

Accenture consultants will be embedded directly into Anthropic’s operations, while Anthropic plans to partner with additional third-party evaluators in the coming weeks to address potential conflicts of interest and help build broader industry consensus.

The embedded personnel will watch models take shape during their training cycles and track the governance decisions that dictate how the systems are built and deployed. They will also conduct direct interviews with Anthropic employees.

The arrangement is one of the first commercial attempts to operationalize independent safety auditing at the training level. Accenture’s auditing services will not be exclusive to Anthropic, as the consulting firm intends to offer its safety evaluation services to other frontier AI developers.

The initiative puts into practice a safety framework proposed by Anthropic Chief Executive Officer Dario Amodei, who has publicly advocated for a slower, more controlled pace of AI development to mitigate systemic risks. His proposed framework outlines a three-step methodology for safe frontier AI development.

At the center of that approach is a requirement that leading AI laboratories grant “ongoing, employee-like access” to external evaluators. Those independent parties are intended to verify that safety protocols are being met, assess model alignment with human values, and report technical incidents or anomalies.

Founded in 2021 by former OpenAI researchers including Dario and Daniela Amodei, Anthropic is based in San Francisco and is known for its Claude family of large language models. The company has long positioned itself as a safety-focused alternative to other major AI labs.

Anthropic previously pioneered “Constitutional AI,” a method of training models to adhere to a set of written principles. It also developed some of the industry’s first “Responsible Scaling Policies,” which call for development to halt if certain risk thresholds are crossed.

Accenture brings experience deploying technology across complex corporate and governmental sectors. The consulting firm announced a separate $3 billion investment in its AI practice in June 2023, and Anthropic hopes Accenture’s knowledge of how enterprises use AI systems in practice will refine its own safety models.

Both companies acknowledge that embedded, real-time AI auditing is entirely experimental. No established industry standards currently govern the physical or digital scope of access evaluators will receive, how proprietary training data will be protected during the audit process, or what formal procedures will dictate the reporting of safety concerns.

The initiative arrives amid intensifying global pressure on AI companies to demonstrate safety compliance. Since the signing of the U.S. Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence in late 2023, federal authorities and international bodies, including the U.S. and U.K. AI Safety Institutes, have sought voluntary and mandatory frameworks to evaluate frontier models before they are released to the public.

The partnership between Anthropic and Accenture represents a private-sector attempt to build an auditing infrastructure before government mandates are fully codified.