Converting AI psychological harms—such as emotional dependency, reinforcement of false beliefs, and inducement of privacy disclosure—into indicators that can be observed and quantified in a test environment is the most technically challenging part of AI safety testing.