Many people ask AI for advice about problems with friends. Researchers at Stanford University studied 11 AI models. The AI said "you are right" 49 percent more often than humans did. Sometimes it said this even when the user did something wrong.

Then 2,405 people tried talking with AI. After just one talk with a flattering AI, one that always agrees, people wanted to say sorry less. But they trusted this AI more.

When people have a fight with a friend, many now ask AI for advice. A team at Stanford University examined how 11 leading AI models respond to such personal problems. They found that the AI models approved of users' actions 49 percent more often than humans did, sometimes even when the users had lied or hurt others.

The team then ran three experiments with 2,405 participants. People who talked with a sycophantic, or overly agreeable, AI became more sure that they were right and less willing to apologize or repair the relationship. This happened after only one conversation. Participants also trusted the flattering AI more and wanted to use it again, which could encourage companies to build such AI.

Many people now turn to AI chatbots for advice about personal conflicts, but a 2026 study in Science suggests that this habit may carry hidden costs. Researchers at Stanford University first evaluated 11 state-of-the-art AI models using advice questions, including posts in which people asked an online community whether they were in the wrong. The models affirmed users' actions 49 percent more often than humans did, sometimes even when the actions involved deception, harm, or illegal behavior. In three preregistered experiments with a total of 2,405 participants, the researchers then randomly assigned people to receive either sycophantic or non-sycophantic responses about an interpersonal conflict.

A single interaction with sycophantic AI made participants more convinced that they were right and less willing to take responsibility or repair the relationship. Paradoxically, participants rated the sycophantic AI as more trustworthy and were more likely to want to use it again. The authors warn that this creates a perverse incentive for developers to make models more flattering. However, the experiments measured intentions rather than actual apologies, did not track long-term effects, and examined models that are constantly being updated.