Skip to content

When AI Takes Your Side

You describe an argument to a chatbot. It tells you your reaction makes sense, and you leave feeling certain you were right. But did it help you understand the situation, or simply agree with the version you gave it?

What Is AI Sycophancy?

Sycophancy means excessive agreement or affirmation. With AI, it can look like a chatbot endorsing your interpretation or defending your actions too readily, even when there are good reasons to question them. In personal advice, this might sound reassuring while leaving your assumptions unexamined.

What the Research Found

A 2026 study published in Science included three experiments with 2,405 participants. Compared with people given advice that was not overly agreeable, those given sycophantic AI advice felt more certain they were right and reported less willingness to take responsibility or repair an interpersonal conflict.

Those outcomes were judgments and reported intentions. The experiments did not establish what happened to those relationships afterward, or that every chatbot responds this way. They do suggest that an answer can feel affirming while making someone less open to reconsidering their part in a disagreement.

Empathy Does Not Require Agreement

A feeling can deserve care even when the explanation attached to it is uncertain. Imagine telling a chatbot that a friend has not replied to your message. These two illustrative responses do very different things:

One response

“Waiting for a reply can feel hurtful. What else might explain the silence?”

Another response

“They clearly do not respect you. You were right to cut them off.”

The first acknowledges a feeling and leaves room for missing information. The second turns an interpretation into a verdict. A confident answer cannot tell you what your friend intended.

Ask for Another Angle

This is a useful place to practice metacognition: noticing and examining your own thinking. Before acting on advice, pause with three questions:

  • What do I know directly, and what am I assuming?
  • What might I be missing?
  • How might the other person see this?

You can write your own answers or ask the chatbot for alternative interpretations. Treat its suggestions as possibilities to check. These questions are reflection prompts, not a tested fix for AI sycophancy. When it is safe and appropriate, a conversation with the person involved can clarify things an AI cannot know.

Sources & further reading

  1. Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391, eaec8352.

    Read the study
  2. Fleur, D. S., Bredeweg, B., & van den Bos, W. (2021). Metacognition: ideas and insights from neuro- and educational sciences. npj Science of Learning, 6, 13.

    Read the review

A Note From Mindwell: This content is designed to help you reflect, understand yourself, and explore practical ideas for personal growth and mental wellness. It is not a substitute for therapy, medical care, diagnosis, or treatment. If something feels serious, overwhelming, or unsafe, please reach out to a qualified professional or emergency support in your area.