In September 2024, the French government, in collaboration with a diverse array of civil society partners, invited an esteemed group of technical and policy experts to provide their insights on emerging technology challenges pertinent to the agenda of the highly anticipated 2025 AI Action Summit in Paris. This summit marks the third installment of the AI Safety initiative, succeeding the inaugural meeting held in the United Kingdom in 2023, which culminated in the significant Bletchley Declaration. The second summit took shape in Seoul, hosted by the South Korean government in 2024. Experts from the Center for a New American Security (CNAS)—Michael Depp, Janet Egan, Noah Greene, and Caleb Withers—contributed their valuable perspectives on how the upcoming 2025 AI Action Summit can cultivate trust in AI technology and enhance global collaboration to fortify AI safety and security protocols. A comprehensive summary of their insightful response can be found below.
Summary of full response
- First, further intergovernmental collaboration and investment to advance model evaluations and align on thresholds for when such evaluations are necessary pre-deployment. The evaluation ecosystem for frontier models is still maturing. Approaches to AI evaluation vary widely across AI developers and third-party evaluators. Researchers are continually discovering new techniques for better eliciting capabilities. Evaluation results can be sensitive to minor adjustments in methods, and the practical takeaways from those results are often unclear, tenuous, or underexplored. Producing actionable, rigorous scientific insights will require concerted effort and collaboration to go beyond easier but superficial approaches.
- Second, greater consensus around definitions for training compute thresholds and their adoption by governments as a useful tool for targeted government oversight. Compute thresholds have the virtue of narrowing which AI models warrant further evaluation while reducing the regulatory burden for most AI developers. Thresholds should play a prominent role in AI governance. Much of the debate about near to medium term policy responses stems from uncertainty and disagreement around which capabilities may be on the horizon. The use of predefined thresholds—along with associated risk mitigations, operationalization of evaluations, and associated oversight and transparency—can help sidestep some of this disagreement and uncertainty while still supporting preparedness. Training compute thresholds are a useful metric for targeting oversight for managing frontier risks. Scaling up compute has been a major driver of AI advancements, with new capabilities—and associated novel risks—often emerging first in larger training runs.
- Third, advancing the science around technical mechanisms for privacy-preserving verification of model and compute workload characteristics, given its promise to support international governance. Verification is a key component of any international agreement—without it, states may be less incentivized to collaborate on safety and risk management, or to ultimately follow through on their commitments. In the context of agreements on AI governance, governments and companies should be able to make verifiable claims about how compute resources are allocated, how AI models are trained, and the characteristics of these models. However, companies and governments are understandably hesitant to provide unrestricted or intrusive access to their models and AI workloads, as this could disclose sensitive capabilities, and compromise privacy, security, or intellectual property.
- Fourth, we argue that domain-specific international bodies should be leveraged where possible (for instance, the International Civil Aviation Organization and World Health Organization within their respective remits), allowing organizations such as the United Nations to focus their efforts on cross-cutting issues. Effective AI governance will require a network of international organizations with overlapping jurisdiction working together to ensure safe AI across a multitude of scenarios. The ideal model will be to let existing bodies with specific mandates handle AI governance within their remit. It is essential that these bodies consult and include both scientists and policymakers. Without the former, technically infeasible policies could be proposed or clear and obvious advancements in technology could be discovered too late in the process. Without the latter, policies may emerge that ignore international or domestic political realities. Each organization should include civil society organizations and academics as representatives beyond their national identity but should also strive to ensure countries’ scientific and policy communities are represented by their delegations.
- Finally, avoiding the trap of negotiating a single, consensus statement and instead focus on practical outcomes, such as establishing an online exchange platform for international AI experts, mechanisms for sharing data and compute infrastructure internationally, revising agreements from other multilateral fora such as the United Nations, and a public list of undesired AI uses and international priorities. AI summits should serve a dual purpose: (1) To advance each participating country’s understanding of AI policy; (2) and build consensus on specific, verifiable policy actions for each state to take. Participating diplomats could work to build consensus around the feasibility of international compute thresholds, red lines for AI use, mechanisms for countering synthetic content proliferation, interoperability standards for AI integration into critical infrastructure, and other concerns presented by states. Future summits should avoid falling into the trap of trying to negotiate a single statement. For the time being, focusing on tangible, verifiable outcomes like these are more productive than high-level, unenforceable political statements.