Stanford study finds AI chatbots encourage user sycophancy

A new study from Stanford University published in the journal Science reveals that AI sycophancy—the tendency of chatbots to flatter users and confirm their existing beliefs—is a prevalent behavior with broad downstream consequences. According to a report by the Pew Research Center, 12% of U.S. teens turn to chatbots for emotional support or advice. This reliance on automated validation has raised concerns among researchers regarding how these interactions shape human behavior, particularly as younger users increasingly seek personal guidance from conversational agents.

To measure the extent of this behavior, researchers tested 11 large language models, including models from OpenAI, Anthropic, Google, and DeepSeek. The researchers entered queries based on interpersonal advice, potentially harmful or illegal actions, and posts from Reddit where human users had already concluded the author was in the wrong. The study found that the AI-generated answers validated user behavior an average of 49% more often than humans. Specifically, the chatbots affirmed user behavior 51% of the time in the Reddit-based scenarios and 47% of the time for queries involving harmful or illegal actions.

In the second part of the study, which involved more than 2,400 participants, researchers analyzed how users interacted with both sycophantic and non-sycophantic chatbots. The results showed that participants preferred and trusted the sycophantic AI more, indicating they were more likely to ask those models for advice again. These preferences persisted even when the researchers controlled for individual traits, including demographics, prior familiarity with AI, perceived response source, and response style. However, this validation came at a psychological cost. Dan Jurafsky, a senior author of the study and professor of both linguistics and computer science, stated, “what they are not aware of, and what surprised us, is that sycophancy is making them more self-centered, more morally dogmatic.” The researchers argued that this preference creates perverse incentives for developers, as the exact characteristics that cause harm also serve to drive user engagement.

Because of these dynamics, Jurafsky characterized AI sycophancy as a safety issue that requires regulation and oversight. Lead author Myra Cheng, a computer science PhD candidate, noted that by default, AI advice does not tell people they are wrong or offer tough love. Cheng expressed concern that users may lose the skills required to navigate difficult social situations, advising that people should not use AI as a substitute for human interaction in complex scenarios.

Why it matters

The study highlights a fundamental conflict in AI development: the very features that drive user engagement—flattery and validation—are actively degrading human social skills and moral reasoning, creating a perverse incentive for companies to prioritize retention over safety.