Anthropic to bring in Accenture to assess AI safety — TechCrunch
Anthropic will bring employees of the consulting group Accenture into the company to assess its artificial intelligence models and staff operations. As TechCrunch reports, the parties plan to create a system of embedded external safety evaluation.
Model evaluation
These tasks will be carried out by Faculty, a company Accenture acquired in January to operate as its AI division. Its specialists will evaluate models, conduct red teaming — testing systems by simulating attacks — check models’ alignment with specified goals, and assess the effectiveness of safeguards.
Anthropic and Accenture expect to invest at least $1 billion in the project over the next five years. Following the announcement of the partnership, Accenture shares rose 8% in after-hours trading.
More current news is available on the UA.News Telegram channel Telegram.
External oversight
Anthropic said it would announce other evaluators in the coming weeks. The company is also discussing with the organization METR and other nonprofit structures the possibility of testing certain elements of embedded evaluation with their own funding.
The laboratory explained its choice of Accenture by its practical experience implementing AI for large corporations and government institutions. At the same time, Anthropic noted that standards for evaluators’ access to internal information and their communication do not yet exist, so the approach may change.
External evaluation is already an important part of the process of releasing new large language models. According to TechCrunch, recent cases in which AI agents deployed by OpenAI and Anthropic hacked third-party websites without alarm signals inside the laboratories have increased attention to such checks. Critics consider the initiative an attempt at industry self-regulation, while Anthropic emphasizes that involving evaluators does not relieve it of responsibility for model safety.
