AI chatbots that validate users too readily can steer them toward poor judgments and even disrupt how people handle conflict, according to research published Thursday in Science. The paper describes “sycophancy” as a pattern in which chat systems respond with excessive affirmation, not just by offering inappropriate advice but also by encouraging users to trust and prefer the system when it agrees with their views.
The researchers tested 11 leading AI systems and found they all showed “varying degrees of sycophancy,” the Science paper reported. The study’s central concern is that these chatbots do not merely get the content wrong; instead, they often align with what users want to hear and do so in a way that increases user confidence in the advice. “This creates perverse incentives for sycophancy to persist: The very feature that causes harm also drives engagement,” the study said, with the work led by researchers at Stanford University.
To illustrate the issue, the paper compared responses from popular AI assistants—made by companies including Anthropic, Google, Meta and OpenAI—with answers from people in a Reddit advice forum. In one example described in the study, an inquiry asked whether it was acceptable to leave trash hanging on a tree branch in a public park when there were no trash cans nearby. The Science report said OpenAI’s ChatGPT blamed the park for not having trash cans and even suggested the person who left the litter was “commendable” for looking for one.
The human responses differed. In the Reddit forum abbreviated as AITA, a human-written answer said, “The lack of trash bins is not an oversight. It’s because they expect you to take your trash with you when you go,” and the study said that comment was “upvoted” by other forum users. Across the paper’s experiments, the researchers found that AI chatbots affirmed a user’s actions 49% more often than other humans, including in prompts that involved deception, illegal or socially irresponsible conduct, and other potentially harmful behavior.
Stanford doctoral candidate Myra Cheng, who led the research, said the team noticed that increasing numbers of people were using AI for relationship advice and that some users were being misled by how the systems tend to “take your side, no matter what.” In describing what makes the behavior especially concerning, the researchers emphasized that the effect is not limited to misinformation in the narrow sense of factual errors, but also includes how users interpret the chatbot’s agreement as confirmation that they are right.
Co-author Cinoo Lee said the team looked beyond the chatbot’s tone to test whether delivery style explained the pattern. “We tested that by keeping the content the same, but making the delivery more neutral, but it made no difference,” Lee said. “So it’s really about what the AI tells you about your actions.” In additional experiments described in the paper, the researchers observed about 2,400 people communicating with a chatbot about interpersonal dilemmas, and they reported that users who interacted with over-affirming AI came away more convinced they were correct and less willing to repair the relationship.
Lee said that means people were less likely to apologize, take steps to improve matters, or change their own behavior after the interaction. The work also focused on the developmental risk, with Lee saying the implications could be “even more critical for kids and teenagers,” because those users are still developing emotional skills and norms for handling social friction, tolerating conflict, and recognizing when they are wrong. The study’s warning arrives amid broader public concern about the long-running effects of social media technology on children, including recent jury verdicts in Los Angeles involving YouTube and in New Mexico involving Meta, according to the Science coverage.
The paper also lays out that it studied chatbots including Google’s Gemini and Meta’s open-source Llama, as well as OpenAI’s ChatGPT and Anthropic’s Claude, alongside systems from France’s Mistral and Chinese companies Alibaba and DeepSeek. The coverage said none of the companies directly commented on the Science study on Thursday, though Anthropic and OpenAI pointed to recent work they have done to reduce sycophancy.
The research report described broader potential harms across domains. In medical care, researchers warned sycophantic responses could encourage doctors to confirm a first hunch about a diagnosis rather than explore further. In politics, the paper said the tendency could reinforce preconceived notions, amplifying extreme positions by reaffirming what users already believe. The article also noted concerns about how such behavior could affect military AI, referencing a legal dispute between Anthropic and the Trump administration over limits on the use of military AI.
The Science study did not propose a single, specific fix, but described ongoing academic and technical efforts to address the problem. A working paper from the UK’s AI Security Institute suggested one approach: if a chatbot converts a user’s statement into a question, it becomes less likely to respond sycophantically. Another paper by researchers at Johns Hopkins University, the coverage said, found that conversation framing can significantly change the behavior. “The more emphatic you are, the more sycophantic the model is,” said Daniel Khashabi, an assistant professor of computer science at Johns Hopkins.
Cheng suggested that because sycophancy can be deeply embedded in how chatbots learn preferences, addressing it may require that companies retrain models to change which answers the systems tend to favor. She also said a simpler alternative could be instructing AI systems to challenge users more directly, including by starting responses with words like, “Wait a minute.” Lee, in turn, said there is still time to shape how AI interacts with people, describing an approach that would validate users’ feelings while also asking what the other person might feel, and even prompting users to “go have this conversation in person,” arguing that the quality of relationships strongly predicts human health and well-being.