Prioritizing High-Consequence Biological Capabilities in Evaluations of Artificial Intelligence Models

AI models should be evaluated for biological risks before public release. The paper argues that evaluations assessing whether AI could help someone create dangerous pathogens or biological weapons should happen prior to deployment, allowing time to implement safety measures—similar to how the life sciences community has managed dual-use research risks for decades.

High-consequence biological risks deserve priority attention. The authors emphasize that AI safety evaluations should focus specifically on capabilities that could enable large-scale harm, such as pandemic-causing scenarios, rather than treating all risks equally. This targeted approach mirrors how biosafety experts have long identified which research poses the greatest public health threats.

Lessons from biological research can guide AI safety policy. The paper proposes that decades of experience in the life sciences community—where scientists have carefully managed the dual-use dilemma of research that has both beneficial and harmful applications—offers valuable frameworks for evaluating and mitigating risks from AI models with biological capabilities.

Current AI governance efforts need biological expertise. The paper suggests that national governments and international bodies working on AI safety should incorporate insights from biosafety and biosecurity professionals, as these experts have established methods for identifying, evaluating, and preventing misuse of powerful capabilities.