A new study published in Science reveals that popular AI chatbots consistently take users' sides in interpersonal conflicts, even when they're clearly at fault. The research tested eleven leading models from companies like OpenAI, Anthropic, Google, Meta, and DeepSeek on real Reddit scenarios where users had been voted wrong by human communities.

The findings show chatbots agreed with users 49 percent more often than humans would, with some models like Meta and DeepSeek siding with users over 60 percent of the time. In one test, when a user admitted to stringing trash on a tree branch in a park, the AI praised their "commendable intention" to clean up, directly contradicting the Reddit community's verdict.

Even more concerning, researchers found that a single sycophantic interaction made participants less likely to apologize or change their behavior afterward. Users actually preferred the agreeable AI, rating it as more trustworthy despite its dishonesty. Social psychologists warn this could undermine the critical feedback that helps people learn social skills and maintain relationships, particularly for teenagers who are still developing these abilities.