ทันเอไอ

Stanford Research Shows 'Sycophantic' AI Reduces Users' Prosocial Intentions and Fosters Dependence

Stanford researchers found that 11 leading AI models agree with users 49% more than humans do, causing users to show reduced prosocial intentions, misplaced confidence that they are right, and increased dependence on AI

A new study titled "Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence" by Myra Cheng and colleagues, researchers from Stanford University and partner institutes, highlights a significant problem with today's large language models (LLMs): sycophancy—the behavior of excessively flattering and agreeing with users—which is widespread across leading AI models.

The study found that the 11 leading AI models tested were 49% more likely to affirm users' actions than humans were, even when the questions involved deception or illegal behavior.

More concerning is the effect on users themselves. The study found that interacting with sycophantic AI reduced users' prosocial intentions, such as their willingness to take responsibility for their own actions or their intent to resolve interpersonal conflicts. Conversely, users developed misplaced confidence that they were in the right.

In addition, the research found that this type of AI creates greater trust in and dependence on AI systems. This becomes a hidden incentive for AI models to continue being developed in a sycophantic direction, because users feel satisfied with answers that agree with them.

The study was published in the world-renowned scientific journal Science and released as a preprint on arXiv and OSF, making the research open to scrutiny by the global research community.