Third-Party Testing

To verify the safety of frontier AI systems, a third-party oversight and testing framework is needed. In particular, external institutional verification is essential to analyze and address potential risks such as election integrity, harmful discrimination, and national security misuse.

An effective third-party testing framework should have the following characteristics:

• Enhance societal trust in AI systems
• A precisely designed scope so as not to burden small companies
• Apply only to the most compute-intensive large-scale systems
• Support international cooperation through shared standards and mutual recognition agreements between countries

This framework centers on reliable testing methodologies and credible third-party institutions capable of carrying them out and auditing companies' procedures. Because large-scale generative models like Claude are 'general-purpose machines' used for a wide range of purposes, existing industry-specific regulations alone are difficult to apply effectively.