The Singapore Consensus on Global AI Safety Research Priorities

Summary

A consensus document from the 2025 Singapore Conference on AI, co-authored by 88 researchers including Bengio, Russell, and Tegmark, organizing AI safety research into a defence-in-depth framework covering Assessment, Development, and Control. It identifies cooperative research priorities where even geopolitical competitors share mutual interest in shared risk thresholds, evaluation protocols, and safety benchmarks. The document builds on the 2025 International AI Safety Report and addresses risks up to and including AGI.

Key Points

• •Organizes AI safety research into three pillars: Assessment (measuring impacts, auditing), Development (specification, training, verification of trustworthy systems), and Control (post-deployment monitoring, societal resilience).
• •Explicitly addresses AGI risk by recommending constraints on autonomy, generality, or intelligence as a development-phase safety strategy.
• •Identifies 'areas of mutual interest' where international competitors can cooperate on shared safety infrastructure like evaluation protocols and benchmarks.
• •Produced alongside ICLR 2025 with 88 co-authors from 11 countries, lending it broad international legitimacy as a coordination document.
• •Covers AI control setups designed to be robust to active subversion, addressing adversarial scenarios in deployed systems.