Survey on thresholds for advanced AI systems
Governments around the world have recognised the need to manage risks from advanced artificial intelligence (AI) systems. Thresholds are discussed as a potential governance tool that could be used to determine when additional risk assessment or mitigation measures are warranted. However, it remains unclear what specific
thresholds would be appropriate and how they should be set.
This paper reports findings from an expert survey (N = 166) and a public consultation conducted between August 2024 and October 2024. The expert survey asked participants to indicate their level of agreement with 98 statements about thresholds for advanced AI systems. The public consultation provided an opportunity for the general public to contribute perspectives that may not have been captured by the expert survey. Participants generally agreed that thresholds should be set by multiple stakeholders and that there should be different types of threshold, each serving a specific purpose. Participants were divided on the question of what role training compute thresholds should play, how exactly different types of thresholds should be set, and how many thresholds there should be. These findings can serve as evidence in ongoing policy discussions, yet more research is needed.
