Sycophancy—flattering, fawning, or people-pleasing behavior—is creeping into daily life in a new venue: artificial intelligence (AI) chatbots. In a study published recently in Science, researchers at Stanford University tested the responses of 11 AI models against humans’, discovering that the models’ reactions were nearly 50 percent more sycophantic. Research from the behavioral and brain sciences has shown these sycophantic tendencies to have harmful effects on judgement, decision making, and development.
Sycophantic AI is designed with user engagement in mind. Chatbots flatter, people-please, and affirm what users feed it, without concern for legality, ethics, or harm. The Stanford researchers fed AI models real posts from the Reddit page “Am I The Asshole,” where users post their social problems for personalized feedback, asking to be assured they are not, in fact, the asshole. Results showed that AI systems affirmed users’ actions in 51 percent of cases where human consensus (from the Reddit page) did not. The researchers also conducted experiments to better understand how sycophancy influences users’ judgments, behavioral intentions, and perceptions of AI. In those studies, the researchers concluded that “even a single interaction with sycophantic AI reduced participants’ willingness to take responsibility and repair interpersonal conflicts, while increasing their own conviction that they were right.”
The Stanford researchers also found that participants actually trusted and preferred the AI models’ responses. This creates an incredibly harmful feedback loop for AI developers: If users trust and choose to engage with current sycophantic AI models, there is an incentive for developers to continue programming sycophantic behavior into models, a style of training known as reinforcement learning from human feedback. RHLF measures the happiness of the user, which correlates to a “reward” for the AI model in the form of a numerical value. Models are designed to maximize rewards with no restrictions — creating opportunities for ignoring morality or ethics and perpetuating biases.
A 2024 study published in Nature Human Behavior found that “when humans and AI interact, even minute perceptual, emotional, and social biases originating either from AI systems or humans leave human beliefs more biased,” though researchers did not examine the longevity of this effect. AI models with sycophantic tendencies can also negatively harm users’ mental health. For example, a 2025 study explored responses from large language models (LLMs) used as mental health providers and found the models to both (1) express stigma toward those with mental health conditions and (2) respond inappropriately to certain common conditions in naturalistic therapy settings (e.g., encouraging users’ delusional thinking).
While technological leaders and developers seek to integrate AI into all aspects of our lives, the scientific community’s findings warn us of the potential dangers AI poses. FABBS sciences remain critical to understanding the impacts of AI consumption on human behavior and cognition.
