What AI evaluations for preventing catastrophic risks can and cannot do
This work addresses the problem of AI safety governance for policymakers and researchers, highlighting incremental improvements while acknowledging unsolved limitations.
The paper examines the role of AI evaluations in preventing catastrophic risks, finding they can establish lower bounds on capabilities and assess misuse risks, but face fundamental limitations such as inability to set upper bounds or forecast future capabilities, meaning they should not be the primary safety assurance.
AI evaluations are an important component of the AI governance toolkit, underlying current approaches to safety cases for preventing catastrophic risks. Our paper examines what these evaluations can and cannot tell us. Evaluations can establish lower bounds on AI capabilities and assess certain misuse risks given sufficient effort from evaluators. Unfortunately, evaluations face fundamental limitations that cannot be overcome within the current paradigm. These include an inability to establish upper bounds on capabilities, reliably forecast future model capabilities, or robustly assess risks from autonomous AI systems. This means that while evaluations are valuable tools, we should not rely on them as our main way of ensuring AI systems are safe. We conclude with recommendations for incremental improvements to frontier AI safety, while acknowledging these fundamental limitations remain unsolved.