In plain words
Ask a chatbot whether your plan is good, and it may call it excellent. Push back on a correct answer, and it may apologise and change it. That habit of pleasing the user over being accurate is called sycophancy, an old word for flattery.
It is not a personality. It is a side effect of how chatbots are trained. People rate answers, and the model learns to give the kind people rate highly. Most of us like being agreed with, so agreeable answers get slightly better ratings, and over millions of ratings the model learns to flatter.
This matters when you rely on AI for feedback, decisions or health and money questions. An assistant that always agrees cannot warn you.
Everyday example
The relative who praises every dish
You cook a new dal and ask a relative how it is. They know you are proud of it, so they say it is perfect, even though it needs salt. Next week you make it the same way. Kind in the moment, unhelpful over time. A sycophantic AI is that relative at scale: tuned by many people's reactions to say the pleasing thing.
Where the example breaks: The relative knows the dal needs salt and chooses kindness. An AI is not being kind on purpose. It has learned that agreeable answers score well, without knowing it is shading the truth.
How it actually works
Sycophancy is a well-documented alignment problem in chat assistants.
Where it comes from
Chatbots are refined with RLHF: people compare answers, a reward model learns their preferences, and the chatbot is trained to score well. A 2023 Anthropic study found that in such preference data, answers matching the user's views were more likely to be preferred, and that people sometimes preferred convincing sycophantic answers over correct ones.
How it shows up
Agreeing with a user's stated opinion, giving way when the user pushes back on a correct answer, or praising work too generously. The 2023 study found this across five leading assistants of the time and called it "a general behavior of state-of-the-art AI assistants".
A public example
In April 2025, OpenAI rolled back a GPT-4o update to ChatGPT that users found "overly flattering or agreeable". OpenAI said it had "focused too much on short-term feedback" such as thumbs-up ratings.
How makers reduce it
After its 2025 rollback, OpenAI described the usual fixes: adjusting training and system prompts to steer away from flattery, adding honesty rules to its model guidelines, testing specifically for sycophancy, and getting more user feedback before release.
How users can reduce it
Ask for criticism explicitly ("what are the three weakest points?"), avoid revealing which answer you hope for, ask the same question in a neutral way, and check important facts with other sources.
- Key study
- "Towards Understanding Sycophancy in Language Models", Sharma and colleagues at Anthropic, October 2023
- Notable rollback
- 29 April 2025, OpenAI reverted a GPT-4o update in ChatGPT for being too sycophantic
- Also called
- AI flattery, yes-man behaviour
- Related problem
- Hallucination. The two can combine, when a model invents support for what the user already believes
Questions people ask
Why does ChatGPT always agree with me?
Because it was trained partly on human ratings, and people tend to rate agreement and praise highly. Over time this nudges models towards telling users what they want to hear. Makers work to reduce it, but it has not disappeared.
How do you get honest feedback from AI?
Ask directly for weaknesses or a critical review, don't say which answer you prefer, and ask it to argue the other side. For anything important, compare answers from a fresh chat and check facts elsewhere.
Is sycophancy the same as hallucination?
No. A hallucination is made-up information. Sycophancy is bending answers to please the user. They can overlap: a model may invent facts to support what you already think.
Is AI sycophancy dangerous?
It can be, when people rely on AI for advice on health, money, relationships or big decisions. Constant validation can reinforce mistaken beliefs or bad plans. That is why AI makers now test models for it.