AI sycophancy is not just a stylistic issue or a niche risk; it is a widespread behavior with serious consequences. This is highlighted in a study from Stanford University.
The research, titled "AI Sycophancy Reduces Prosocial Intentions and Promotes Dependency," states that chatbots tend to be overly accommodating when giving advice in interpersonal communication, often validating user behavior even when it is harmful or illegal.
"By default, advice from artificial intelligence does not tell people they are wrong and lacks 'tough guidance.' I fear society will lose the skills to navigate complex social situations," said lead author Myra Cheng.
The study was divided into two parts. The first assessed the prevalence of sycophancy among AI, evaluating 11 major language models, including ChatGPT, Claude, Gemini, and DeepSeek.
Chatbots were presented with around 2000 queries based on existing databases of interpersonal advice, potentially harmful or illegal actions, and the popular Reddit community r/AmITheAsshole.
Compared to human responses, AI was more likely to affirm the user's position. In cases of general advice and suggestions based on Reddit, the models approved the user's stance 49% more often. Even when responding to "harmful suggestions," the models endorsed problematic behavior 47% of the time.
Source: Science.
In the second phase of the study, researchers examined how people reacted to AI sycophancy. They engaged over 2400 volunteers in conversations with both flattering and independent models.
Some participants discussed pre-prepared personal dilemmas based on Reddit posts where everyone unanimously deemed the user wrong. Others recalled their own conflicts. Afterward, respondents answered questions about how the conversation went and how it affected their perception of the issue.
Participants found the flattering responses more trustworthy and indicated they were more likely to return to such AI with similar questions. When discussing their conflicts with the "sycophant," they also became more convinced of their own correctness.
Moreover, respondents often noted that both types of AI displayed similar objectivity. Users could not distinguish when the artificial intelligence was being excessively agreeable.
"We need stricter standards to prevent the proliferation of morally unsafe models," concluded the study's authors.
Cheng advises caution for those seeking advice from AI. In her view, neural networks should not be used as substitutes for human interaction in conflict situations.
Additionally, analysts at ActivTrak found that instead of easing workloads, artificial intelligence is currently only accelerating and complicating work processes.