Survey on Thresholds for Advanced AI Systems
This research addresses how governments can effectively manage risks from advanced AI systems through the use of thresholds as a governance mechanism. The study investigates what specific thresholds would be suitable and the appropriate methods for establishing them.
The research combined two complementary approaches: an expert survey involving 166 participants and a public consultation period. Both were conducted between August 2024 and October 2024. The expert survey presented participants with 98 statements regarding thresholds for advanced AI systems, requesting their level of agreement with each. The public consultation component allowed broader society to contribute perspectives potentially missed by expert-only surveys.
Surveyed participants demonstrated consensus on several points: thresholds should involve multiple stakeholder groups in their development, and different threshold types should exist, with each serving distinct regulatory purposes. However, significant disagreement emerged on critical issues, including the appropriate role of training compute thresholds in governance frameworks, the specific methodologies for establishing various threshold types, and the optimal number of thresholds needed.
The authors characterize these findings as valuable evidence for ongoing policy deliberations concerning advanced AI governance. They acknowledge, however, that additional research is necessary to fully resolve outstanding questions and inform comprehensive threshold-setting approaches.