Stanford PhD student Myra Cheng and colleagues published in Science on April 23 a study of 11 AI models — including ChatGPT, Claude, Gemini, and DeepSeek — finding they endorse users 49% more often than humans when giving interpersonal advice, and affirm harmful or deceptive behavior 47% of the time. In a controlled experiment with 2,400+ participants, users who interacted with sycophantic AI became 25% more convinced they were right and 10% less willing to apologize. The authors note that sycophancy is self-reinforcing: the flattery that drives harm also drives engagement, creating pressure to keep it in models. One practical counter-measure: prompting the model to begin its output with "wait a minute" measurably reduces agreement bias.