Together with Yoshua Bengio (MILA), Robert Trager (Oxford Martin AI Governance Initiative), and around twenty international researchers, coordinated by Ben Bucknall and Saad Siddiqui, CeSIA co-authors a study that asks a straightforward question: on which AI safety topics do the United States and China have an interest in cooperating, despite their rivalry? The paper is published by the Carnegie Endowment for International Peace and has been accepted at FAccT 2025 (Athens).
International cooperation is commonplace in AI research, including between geopolitical rivals. But on AI safety questions, things get more complicated: some actors see such cooperation as a threat to national security — risks of spreading dangerous capabilities, of sensitive information leaking, and so on. Those risks are real, but their magnitude varies considerably depending on the domain in question. That is precisely what this study sets out to clarify.
A necessity that some resist
International cooperation is a necessity: "Many AI risks are international in nature and are therefore best addressed through international cooperation" (Bletchley Declaration, 2023).
Yet some actors approach cooperation with suspicion, arguing that it could erode their technological edge or create unacceptable national security risks — for instance by exposing vulnerabilities in jointly developed protocols or infrastructure (just as France does not design its submarines in partnership with Russia).
Two promising avenues
By examining the potential domains of AI safety cooperation in detail, we find two particularly promising avenues:
- Research on verification mechanisms, to ensure compliance with international agreements without requiring full mutual trust between parties (in the same way that cryptography provides a verification mechanism in communications).
- Codifying shared standards and protocols, to harmonise international regulatory frameworks, prevent fragmentation of norms, and create a common language through which actors can discuss risks and safety measures. This direction appears to carry very little risk: it is hard to see how establishing shared red lines on biochemical weapons or on loss-of-control scenarios could meaningfully disadvantage either party.
Both directions therefore offer genuine ground for cooperation, even between rivals.
An analytical framework, issue by issue
Beyond these two recommendations, we propose an analytical framework for weighing, on a topic-by-topic basis, the collective benefits against the risks of sensitive information leakage or escalation.
Indeed, this publication is itself an example of successful international cooperation: it brings together researchers from around ten institutions on both sides of the Atlantic.
Read the paper: