Formal Definition of Catastrophic Performance in AI Misalignment
November 5, 2026
This paper introduces a formal framework to distinguish between suboptimal AI behavior and catastrophic misalignment. It uses baseline and orthogonal reference values to quantify risks where an agent's trajectory deviates fundamentally from human preference trajectories.
HOW THIS AFFECTS YOU
●
researcherYou can use these two reference values to conduct more precise safety and alignment evaluations.
●
policyThis provides a more rigorous mathematical basis for defining and regulating catastrophic AI risks.