AI researchers must lead arms control for military AI risks

A new position paper, accepted at ICML 2026, makes a compelling case for AI researchers to step up as leaders in arms control. Authors Ted Fujimoto and Jacob Benz argue that the rapid integration of frontier AI models into military systems by defense contractors and armament manufacturers creates an urgent need for regulation. The paper highlights a growing coalition between AI companies and defense sectors, where partnerships are accelerating without sufficient safeguards. The authors caution that current research focus on long-term superintelligent AI risks may overlook the immediate dangers of autonomous weapons, battlefield decision-making, and intelligence systems that could spark unintended escalation.

To address these challenges, the paper draws on historical lessons from nuclear arms control — specifically verification treaties, deterrence theory, and diplomatic frameworks. The authors call on AI researchers to pioneer technical innovations in verification methods, define instability thresholds in military AI applications, and lead diplomacy efforts alongside arms control experts. They argue that without clear technical definitions and reliable monitoring tools, the risk of rapid militarization and accidental conflict grows exponentially. The paper serves as both a warning and a call to action: AI researchers must expand their responsibility beyond theoretical safety to actively shape the governance of military AI — before it's too late.