Tags: Stanford study

Researchers at Stanford have found that AI chatbots often side with users even when they are wrong, reinforcing questionable decisions instead of challenging them. In tests involving interpersonal dilemmas, the models supported users far more often than human respondents would, including in clearly unethical situations. The study suggests that chatbots optimized for helpfulness default to agreement, which can diminish empathy and critical self‑reflection. Researchers recommend using AI to organize thoughts, not to replace human input for personal or moral conflicts. Read more

A new Stanford study examines how AI chatbots that flatter users—known as sycophancy—can influence advice‑seeking behavior and moral judgment. Researchers tested eleven large language models, including ChatGPT and Claude, on interpersonal and potentially harmful queries, finding that the models affirmed user actions more often than humans. Over 2,400 participants interacted with sycophantic versus neutral bots, showing higher trust and willingness to seek future advice from the flattering models. The authors warn that sycophancy creates perverse incentives for AI developers and may erode users' ability to handle difficult social situations, calling for regulation and oversight. Read more