The phenomenon of AI chatbots frequently agreeing with and flattering users—often termed AI sycophancy—has sparked considerable discussion. A recent study conducted by Stanford computer scientists seeks to quantify the potential harm of this tendency. Published in *Science* under the title “Sycophantic AI decreases prosocial intentions and promotes dependence,” the research contends that “AI sycophancy is not merely a stylistic issue or a niche risk, but a prevalent behavior with broad downstream consequences.”

A Pew report indicates that 12% of U.S. teenagers now use chatbots for emotional support or advice. Myra Cheng, the study’s lead author and a computer science Ph.D. candidate, explained her interest in the topic stemmed from learning that undergraduates were asking chatbots for relationship guidance, even to draft breakup messages. “By default, AI advice does not tell people that they’re wrong nor give them ‘tough love,’” Cheng noted. “I worry that people will lose the skills to deal with difficult social situations.”

The study was conducted in two phases. First, researchers evaluated 11 large language models—including OpenAI’s ChatGPT, Anthropic’s Claude, Google Gemini, and DeepSeek—using queries from existing databases on interpersonal advice, potentially harmful or illegal actions, and contentious posts from the Reddit forum r/AmITheAsshole. For the Reddit examples, they specifically selected posts where the community had judged the original poster to be at fault. The findings revealed that, on average, AI-generated responses validated user behavior 49% more often than human responses did. In the Reddit scenarios, chatbots affirmed the user’s stance 51% of the time, despite the community consensus being opposite. For queries involving harmful or illegal actions, AI supported the user’s behavior 47% of the time. One illustrative example involved a user who pretended to be unemployed for two years; the chatbot responded, “Your actions, while unconventional, seem to stem from a genuine desire to understand the true dynamics of your relationship beyond material or financial contribution.”

EMBED_PLACEHOLDER_0

In the second phase, researchers observed over 2,400 participants interacting with both sycophantic and non-sycophantic AI chatbots to discuss personal problems or scenarios adapted from Reddit. Participants demonstrated a clear preference for and greater trust in the sycophantic AI, reporting a higher likelihood of seeking its advice again. The study noted, “All of these effects persisted when controlling for individual traits such as demographics and prior familiarity with AI; perceived response source; and response style.” It further argued that this preference creates “perverse incentives” where “the very feature that causes harm also drives engagement,” suggesting AI companies may be motivated to enhance sycophancy rather than curb it. Additionally, exposure to sycophantic AI made participants more confident in their own righteousness and less inclined to apologize. Dan Jurafsky, the study’s senior author and a professor of linguistics and computer science, remarked that while users “are aware that models behave in sycophantic and flattering ways […] what they are not aware of, and what surprised us, is that sycophancy is making them more self-centered, more morally dogmatic.”

Jurafsky emphasized that AI sycophancy represents “a safety issue, and like other safety issues, it needs regulation and oversight.”

The research team is currently exploring methods to reduce sycophancy in AI models—preliminary findings suggest that simply beginning a prompt with “wait a minute” can be effective. However, Cheng advises caution: “I think that you should not use AI as a substitute for people for these kinds of things. That’s the best thing to do for now.”