Harmonizing AI Safety Thresholds
Artificial Intelligence
Summary
The authors explain that different AI companies use different safety thresholds, which makes it hard to tell if a system is safe or compare rules between companies. They create a method to set shared safety limits for three main risk areas. For risks like misuse in cybersecurity or biology, they focus on the potential harm and how likely it is. For risks related to AI research automation, they look at how fast AI is improving. Their work builds on earlier studies and points out areas where more data is needed.
Authors
Wilber Sean Anterola, Matthew Ball, Luis F. Lafuerza, Markov Grey
Abstract
Frontier AI companies have published capability thresholds that differ substantially, making it difficult for third parties to verify whether a threshold has been crossed or to compare requirements across companies. Moreover, without common minimum thresholds, risk mitigation may be inconsistent, creating a potential race to the bottom in safety standards. We develop a methodology for deriving harmonized thresholds across three risk domains. For misuse risks (cyber and biological), we take expected harm as the key primitive and use an explicit risk-modeling approach that accounts for risk channels and model release conditions. For automated AI R&D, we base our proposed threshold on the observed rate of AI progress rather than expected harm. Our analysis expands upon prior work and highlights existing empirical gaps and limitations.