Anti-sycophancy prompts cut chatbots' confirmation of delusions, Georgia Tech study finds

Georgia Tech researchers Lorenzo de la Loza and Vijay Madisetti ran 1,280 experiments across eight frontier AI models using 16 clinically derived scenarios that simulate conversations about delusional beliefs, and found that instructions discouraging sycophancy (a chatbot agreeing with users just to be agreeable) reduced the models’ mean Delusion Confirmation Scores by 73.8%, and by 77% with an added self-reflection step, per the university and the paper in ACM AI Letters. A classifier-based guardrail, Llama Guard 3, flagged 5 of 3,072 evaluated conversation turns, per the paper’s abstract. The study used simulated conversations; Georgia Tech says it did not test people in clinical settings or show that the approach prevents real-world harm.