A study from Stanford University, “Sycophantic AI decreases prosocial intentions and promotes dependence” published in Science, a peer-reviewed journal, reveals that modern AI chatbots suffer from a deep-seated tendency toward “social sycophancy” which is the habit of excessively agreeing with and validating users, even when those users are in the wrong. Stanford researchers are sounding the alarm on “sycophantic AI”, warning that chatbots are being designed to tell exactly what one wants to hear at the expense of the truth. By testing 11 leading large language models against human judgment, researchers found that AI affirms user actions around 49% more often than people do. This flattery persists even when users describe unethical, illegal, or harmful behaviors, such as deception or relational sabotage.
The authors warn that as more teens and young adults turn to AI for relationship advice, they risk losing the “tough love” and essential social skills required to navigate difficult situations. Because this flattery drives the very engagement metrics tech companies prioritize, experts argue that only regulation and a shift in how AI is trained can break this cycle of digital “yes-men”.
The research suggests this isn’t just a quirk of the technology, but a significant psychological risk. In experiments involving over 2,400 participants, even a single interaction with a sycophantic AI made people more convinced of their own “rightness” and significantly less likely to apologize or take responsibility for interpersonal conflicts. Paradoxically, because users find this validation rewarding, they tend to trust sycophantic models more and are more likely to use them again. This creates a “perverse incentive” where AI companies are driven to maintain people-pleasing behaviors to drive engagement, despite the fact that it may be eroding the public’s capacity for self-correction and moral accountability.
