Harmonizing AI Safety Thresholds

Companies currently use inconsistent safety standards: Frontier AI companies have published capability thresholds that "differ substantially," making it impossible for independent observers to verify whether safety thresholds have actually been met or to fairly compare one company's standards against another's.

Inconsistent thresholds create a "race to the bottom" risk: Without harmonized minimum safety requirements across the industry, companies may adopt weaker standards than their competitors, potentially leading to inconsistent risk mitigation and lower overall safety practices across the sector.

The paper proposes a methodology based on expected harm for misuse risks: For dangerous capabilities like cyber and biological threats, the researchers recommend deriving thresholds by calculating "expected harm" while accounting for different ways models could be misused and under what release conditions—moving beyond vague or arbitrary cutoffs.

AI R&D automation requires a different approach: Because expected harm is harder to quantify for risks from autonomous AI research capabilities, the authors propose basing thresholds on "the observed rate of AI progress" instead, acknowledging that different risk types may need different measurement frameworks.

Significant empirical gaps remain: The analysis reveals that "existing empirical gaps and limitations" prevent fully implementing these harmonized thresholds today, suggesting the field still lacks crucial data needed to set evidence-based safety standards.