Advancing AI Safety: Key Insights for the 2025 AI Action Summit in Paris

Advancing AI Safety: Key Insights for the 2025 AI Action Summit in Paris

November 27, 2024 Catherine Williams - Chief Editor World

In September 2024, the French government worked with civil society partners to invite experts to discuss emerging technology topics relevant to the 2025 AI Action Summit. This summit is the third iteration of the AI Safety Summit series. The first summit was held in the UK in 2023, leading to the Bletchley Declaration. South Korea hosted the second summit in 2024. Experts from CNAS offered insights on enhancing trust in AI and promoting global cooperation for AI safety.

Summary of Key Points:

  1. Collaboration on Model Evaluations: Governments should collaborate and invest in model evaluations to set clear thresholds for pre-deployment checks. The evaluation process for advanced AI models is still developing. Diverse approaches exist among AI developers, creating uncertainty. To produce clear insights, stakeholders must work together and avoid superficial methods.

  2. Consensus on Training Compute Thresholds: Governments should agree on training compute thresholds to streamline oversight. These thresholds help identify which AI models need thorough evaluation and reduce the regulatory burden on most developers. By using predefined thresholds, governments can manage risks more effectively without adding unnecessary complexity.

  3. Privacy-Preserving Verification: Advancing verification techniques for AI models is vital for international collaboration. Verification ensures that countries can trust each other’s claims regarding AI training and computing resources. However, transparency must be balanced with the need to protect sensitive information and intellectual property.

  4. Leveraging Domain-Specific Organizations: Existing international bodies like the International Civil Aviation Organization and the World Health Organization should handle AI governance within their areas. This approach allows broader organizations, such as the United Nations, to address cross-cutting issues. It’s essential that these bodies include scientists and policymakers to create feasible and effective policies.

  5. Focus on Practical Outcomes: AI summits should prioritize practical results instead of aiming for a single consensus statement. Summits can aim to create an online platform for sharing information, mechanisms for data exchange, and guidelines for AI use. Participants should target tangible actions that improve understanding of AI policy and ensure each state takes specific steps to address AI challenges.

By addressing these points, the 2025 AI Action Summit can enhance cooperation and safety in AI development worldwide.