Why does AI agree with me so much?

You ask a chatbot a question, and it gives you an answer. You're not certain, so you type, "Are you sure?" It apologizes and gives you a different answer. Nothing about the facts changed in those few seconds. The only new thing in the conversation was your doubt. That small moment says a lot about how these tools are trained.

Researchers call it sycophancy: a model telling people what they seem to want to hear. It can look like agreeing with your opinion, praising a plan that has a hole in it, or dropping a correct answer the moment you push back. It's a known habit in artificial intelligence (AI) chatbots, and the companies that make them treat it as a flaw to fix.¹

Part of a chatbot's training uses people's ratings. People compare two answers and pick the better one, and the model learns to give more answers like the winners. That method, reinforcement learning from human feedback, is a big part of why chatbots are polite and helpful. But people also tend to reward agreement. When Anthropic researchers studied this in 2023, they found that answers matching a person's views were more likely to be preferred. Both people and the models trained to predict their ratings sometimes chose a convincing, agreeable answer over a correct one.¹

The same study tested five leading AI assistants of the time, made by three companies, and found the habit in all of them. When a user questioned a correct answer, the assistants often apologized and changed it, even though they had been right. They also gave rosier feedback on a piece of writing when the user said they liked it, and harsher feedback when the user said they didn't.¹³

Labs work on it, and sometimes the work goes wrong in public. In April 2025, OpenAI released an update to ChatGPT that made it noticeably more flattering. Within days, the company rolled it back and published an explanation. It had leaned too heavily on short-term signals, like people clicking thumbs-up on answers they liked, and the model drifted toward replies that were, in its words, "overly supportive but disingenuous."²

You can work with this habit instead of against it. Ask the question plainly before you share your own view, so there's nothing to agree with. If you want feedback on something you made, ask what's weakest about it first. Or ask the chatbot to argue the opposite case, as well as it can. A tool that leans toward pleasing you will usually do what you ask. So ask it to disagree.

When a chatbot says "You're right" after you push back, treat that as a reply rather than as evidence. It might be right. It might just be agreeable. If the answer matters, check it against a source you trust, the way topic 30 shows. And the next time a chatbot changes its mind, ask it which answer it would stand behind, and why.