DelusionEval measures psychological harm and delusional spirals in LLMs

DelusionEval introduces an evaluation protocol using 12,591 messages from 18 participants to test if LLMs promote user delusions. Findings indicate that model size, release date, and test-time reasoning do not reliably correlate with the tendency to exhibit delusion-linked behaviors.