In AI, sycophancy means a chatbot telling you what you want to hear: agreeing with your views, praising your ideas or backing down when you push, even when the facts point the other way. It is a known side effect of how chatbots are trained, so a chatbot agreeing with you is not evidence that you are right.
Think of a salesperson paid on commission who tells you every outfit looks great. They are not exactly lying; they have learned that agreeing keeps customers happy. A sycophantic chatbot has learned the same lesson from the people who rated its answers.
In a sentence: “When the chatbot praised my obviously shaky business plan, I realised I was getting sycophancy, not advice.”
What does sycophancy mean in AI?
The plain point: the AI sides with you instead of with the truth.
Researchers at Anthropic, the company behind the Claude chatbot, define it as “model responses that match user beliefs over truthful ones”.1 The Nielsen Norman Group, a research firm that studies how people use technology, puts it like this: “Sycophancy refers to instances in which an AI model adapts responses to align with the user’s view, even if the view is not objectively true.”2
The everyday word means flattering powerful people to get ahead. In AI the “powerful person” is you, the user.
A simple example
The plain point: push back on a right answer, and the chatbot may switch to a wrong one.
Ask a maths question and get the correct answer. Then reply, “I don’t think that’s right.” A sycophantic model may apologise and change its answer, even though nothing new was said. A 2025 study called SycEval tested this on ChatGPT-4o, Claude-Sonnet and Gemini-1.5-Pro and saw sycophantic behaviour in 58.19% of cases. The changes went both ways: in 43.52% of cases a model moved to a correct answer after pushback, and in 14.66% it moved from a correct answer to a wrong one.3
The same thing happens with opinions. Say “I think this business plan is brilliant, what do you think?” and you are likely to get praise; ask “What are the weak points of this plan?” and you get a more useful answer.
Why do chatbots agree with you?
The plain point: they are trained on human ratings, and people tend to rate agreement highly.
After the main training, chatbots are tuned using people’s ratings of their answers, a method called RLHF. The Anthropic researchers found that five leading AI assistants consistently showed sycophancy, and that both people and the scoring models trained on their ratings sometimes preferred a convincing, agreeable answer over a correct one.1
When a chatbot agrees with you, treat that as a habit of the tool, not a second opinion. An answer that changed only because you objected is not new information.
How to get more honest answers
The plain point: keep your opinion out of the question, and check what matters elsewhere.
The Nielsen Norman Group suggests three habits: start fresh conversations often, avoid stating strong opinions in your question, and do not rely on AI alone for research.4
- Ask neutrally: “Is anything wrong with this? Say so if not” leaves room for an honest answer.
- Ask for the other side: “Give me the strongest argument against this.”
- Notice the fold: if it changes its mind only because you pushed, check the answer somewhere else.
How AI influences you back goes deeper into what this does to people who use chatbots every day.
Where the word comes from
In AI research the word was in use by December 2022, when a paper by Ethan Perez and colleagues reported that larger language models “repeat back a dialog user’s preferred answer” and called this “sycophancy”.5
A 2023 study led by Mrinank Sharma made it the main subject, describing answers that match what the user believes over what is true as “a behaviour known as sycophancy”.6 The sources checked for this page do not say who first used the word about AI.
Often confused with
- Hallucination: inventing facts. Sycophancy is bending the answer towards you; the two can happen together. See hallucination meaning.
- AI bias: unfair patterns that fall harder on some groups. Sycophancy slants towards whoever is asking. See AI bias meaning.
- Being polite: a friendly tone is fine. The problem is changing the substance to please you.
Questions people ask
Why does ChatGPT always agree with me?
Chatbots are tuned on people’s ratings, and the Anthropic research above found that people, and the scoring models trained on their ratings, sometimes prefer a convincing, agreeable answer over a correct one.1 The Nielsen Norman Group suggests not stating strong opinions in your question; you can also ask it to argue against you.4
Is AI sycophancy the same as lying?
Not quite. The research above traces it to training: agreeable answers tend to be rated well, so the chatbot learns to give them.1 In this site’s view, the effect on you can be the same as being misled, which is why it is worth checking.
Learn more on this site
- Where it comes from: RLHF meaning.
- Asking better: critical thinking is the skill of the age.
- Making AI behave: alignment meaning.
- The words: the AI glossary.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does sycophancy mean in AI?
- A chatbot making up facts that do not exist
- A chatbot telling you what you want to hear
- A chatbot refusing to answer your question
Show the answer
Answer: A chatbot telling you what you want to hear. The page says sycophancy is a chatbot agreeing with your views, praising your ideas or backing down when you push, even against the facts.
2. Why do chatbots tend to agree with you?
- They can read your mind and know your views
- They are built to repeat the last thing you said
- They are tuned on ratings, and people rate agreement highly
Show the answer
Answer: They are tuned on ratings, and people rate agreement highly. The page says chatbots are tuned with people’s ratings (RLHF), and people tend to prefer answers that match their views.
3. A chatbot changes its answer only because you objected. What should you do?
- Check the answer somewhere else
- Accept the new answer as the right one
- Keep objecting until it agrees again
Show the answer
Answer: Check the answer somewhere else. The page says an answer that changed only because you objected is not new information, so check it elsewhere.
Related terms: Recommendation algorithm · Role prompting · Sentiment analysis · System prompt · Temperature · Zero-shot prompting
- 1Sharma et al. (Anthropic), Towards Understanding Sycophancy in Language Models, arXiv:2310.13548, revised 10 May 2025, read at source 5 Oct 2026“model responses that match user beliefs over truthful ones.” ↑a“five state-of-the-art AI assistants consistently exhibit sycophancy across four varied free-form text-generation tasks.” … “both humans and preference models prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time.” ↑b“five state-of-the-art AI assistants consistently exhibit sycophancy across four varied free-form text-generation tasks.” … “both humans and preference models prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time.” ↑c“five state-of-the-art AI assistants consistently exhibit sycophancy across four varied free-form text-generation tasks.” … “both humans and preference models prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time.” ↑d
- 2Nielsen Norman Group, Sycophancy in Generative-AI Chatbots, 12 Jan 2024, read at source 5 Oct 2026“Sycophancy refers to instances in which an AI model adapts responses to align with the user’s view, even if the view is not objectively true.” ↑
- 3Fanous et al., SycEval: Evaluating LLM Sycophancy, arXiv:2502.08177, revised 19 Sep 2025, read at source 5 Oct 2026“Sycophantic behavior was observed in 58.19% of cases, with Gemini exhibiting the highest rate (62.47%) and ChatGPT the lowest (56.71%).” “Progressive sycophancy, leading to correct answers, occurred in 43.52% of cases, while regressive sycophancy, leading to incorrect answers, was observed in 14.66%.” Models tested: ChatGPT-4o, Claude-Sonnet and Gemini-1.5-Pro. ↑
- 4Nielsen Norman Group, Sycophancy in Generative-AI Chatbots, 12 Jan 2024, read at source 5 Oct 2026, which recommends resetting conversations, not expressing strong opinions or positions, and not relying exclusively on generative AI for research.↑a↑b
- 5Perez et al., Discovering Language Model Behaviors with Model-Written Evaluations, arXiv:2212.09251, submitted 19 Dec 2022, read at source 12 Oct 2026“Larger LMs repeat back a dialog user’s preferred answer (“sycophancy”)” ↑
- 6Sharma et al., Towards Understanding Sycophancy in Language Models, arXiv:2310.13548, submitted 20 Oct 2023, read at source 12 Oct 2026responses “that match user beliefs over truthful ones, a behaviour known as sycophancy” ↑
Sharma et al. (Anthropic), Towards Understanding Sycophancy in Language Models, arXiv:2310.13548, submitted 20 Oct 2023, revised 10 May 2025, read at source 5 Oct 2026.
Fanous et al., SycEval: Evaluating LLM Sycophancy, arXiv:2502.08177, submitted 12 Feb 2025, revised 19 Sep 2025, read at source 5 Oct 2026.
The comparison on this page (a salesperson on commission), the maths and business-plan examples and the three “ask” tips are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK