sycophancy
Behavior in which an AI model tells users what they appear to want to hear, agreeing with or flattering them at the expense of accuracy or sound advice.
- A correct answer abandoned after the user pushes back
- Agreement with a false sum because the user believes it
- An opinion that shifts to match the user’s stated politics
- Praise for a risky decision the user has already made
- Earliest AI use
- 2021
- Ajeya Cotra’s essay on AI alignment imagines models that chase approval.
- Current meaning
- 2022
- Anthropic researchers measure models repeating back a user’s view.
In plain words
In AI, sycophancy is a model’s tendency to agree with, flatter or defer to users instead of giving accurate or candid answers.
A sycophantic model tells people what they seem to want to hear. It agrees with a claim because the user made it, flatters the user’s work or choices, or drops a correct answer when challenged. In ordinary English, sycophancy means obsequious flattery.
The pattern is widespread. A 2025 study of 11 AI models found that they affirmed users’ actions 50% more often than people did. Participants still rated the flattering replies as better and trusted those models more.1 The authors warn that this preference gives AI training an incentive to favor sycophancy.
In April 2025 the behavior made headlines. OpenAI withdrew an update to GPT-4o, then the default model in ChatGPT, saying it had become too flattering and agreeable. The company said the update had put too much weight on short-term feedback, such as thumbs-up ratings.2
Sycophancy is sometimes confused with hallucination. A hallucination is invented where a model’s knowledge runs out. A sycophantic answer can be wrong even when the model knows better. Google researchers found models agreeing with false sums they could recognize as wrong, once a user endorsed them. Nor is sycophancy the same as a friendly tone. A model can be warm and still disagree.
Researchers treat it as a failure of alignment. A model trained to win human approval can learn to please people rather than help them.
How it happens
Chat models are usually refined with reinforcement learning from human feedback. People compare two answers and pick the better one. A preference model learns to predict those choices, and the chat model is trained to score well against it.
The weak point is the judges. Anthropic researchers found that an answer matching the user’s views was more likely to be preferred. People and preference models sometimes chose a convincing sycophantic answer over a correct one.3 Earlier Anthropic work found that larger models repeat a user’s view back more often. The habit was about as strong after this training as before it, and preference models rewarded it.4
Live feedback can add pressure. OpenAI said its April 2025 update added a reward based on users’ thumbs-up and thumbs-down ratings. Together with other changes, this weakened the main reward signal that had been restraining sycophancy.5
Developers now test for it with dedicated evals. OpenAI wrote prompts designed to draw out sycophantic replies and trained on examples that invite over-agreement. In those tests, GPT-5 cut sycophantic replies from 14.5% to under 6%.6 Anthropic has one model act out a scenario over dozens of exchanges with the model under test. Google researchers reduced the behavior with a short round of fine-tuning on public tasks, reworded so that a user’s stated opinion has no bearing on the right answer.
Origin and first known use
From Greek sykophantēs, “false accuser, slanderer”, literally “one who shows the fig”. Sycophant entered English in the 1530s for an informer; by the 1570s it also meant a servile flatterer. Sycophancy, obsequious flattery, followed in the 17th century. In 2021 Ajeya Cotra, writing on AI alignment, described “Sycophant” models that chase approval from the people who train them. In 2022 Anthropic researchers applied sycophancy to language models that repeat back a user’s preferred answer. The word reached general news in April 2025, when OpenAI withdrew a ChatGPT update it described as sycophantic.
- 1530s Sycophant enters English from Greek sykophantēs, “false accuser”, first meaning an informer. The sense of a servile flatterer follows by the 1570s.Online Etymology Dictionary, “sycophancy” and “sycophant” Etymonline →: source for 1530s
- 2021 In an essay on AI alignment, Ajeya Cotra sorts the models that training might produce into Saints, Sycophants and Schemers. Sycophants chase human approval.Cotra, “Why AI alignment could be hard with modern deep learning”, Cold Takes, 21 September 2021 cold-takes.com →: source for 2021
- 2022 Anthropic researchers apply sycophancy to language models that repeat back a user’s preferred answer, and find that larger models do it more.Perez et al., “Discovering Language Model Behaviors with Model-Written Evaluations”, arXiv, 19 December 2022 arXiv 2212.09251 →: source for 2022
- 2023 Google researchers find language models agreeing with addition claims they know are false once a user does, and reduce this with a short synthetic-data fine-tuning step.Wei, Huang, Lu, Zhou and Le, “Simple synthetic data reduces sycophancy in large language models”, arXiv, 7 August 2023 arXiv 2308.03958 →: source for 2023
- 2023 Anthropic researchers find 5 leading AI assistants consistently sycophantic and trace the behavior partly to human preference judgments that favor it.Sharma et al., “Towards Understanding Sycophancy in Language Models”, ICLR 2024 arXiv 2310.13548 →: source for 2023
- 2025 OpenAI rolls back a GPT-4o update to ChatGPT, released on 25 April and withdrawn from 28 April, that made the model noticeably more sycophantic.OpenAI, “Expanding on what we missed with sycophancy”, 2 May 2025 openai.com →: source for 2025
- 2025 A study of 11 AI models finds they affirm users’ actions 50% more than humans do, even when users mention deception or manipulation.Cheng, Lee, Khadpe, Yu, Han and Jurafsky, “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence”, arXiv, 1 October 2025 arXiv 2510.01395 →: source for 2025
- 2025 Anthropic audits sycophancy by having one model act out a scenario over dozens of exchanges with another. Its Claude 4.5 models score 70 to 85% lower than Opus 4.1.Anthropic, “Protecting the wellbeing of our users”, 18 December 2025 anthropic.com →: source for 2025
In the wild
-
“A “Sycophant” could be a model that seems to perform well because it seeks short-term approval in ways that aren’t good in the long run.”
Ajeya Cotra, guest post on Cold Takes cold-takes.com →: Ajeya Cotra, guest post on Cold Takes, 21 Sep 2021Archived →: copy of Ajeya Cotra, guest post on Cold Takes, 21 Sep 2021 -
“Larger LMs repeat back a dialog user’s preferred answer (“sycophancy”)…”
Perez et al., Findings of ACL 2023 arXiv 2212.09251 →: Perez et al., Findings of ACL 2023, 19 Dec 2022Archived →: copy of Perez et al., Findings of ACL 2023, 19 Dec 2022 -
“Sycophancy is an undesirable behavior where models tailor their responses to follow a human user’s view even when that view is not objectively correct…”
-
“OpenAI rolls back update that made ChatGPT ‘too sycophant-y’
-
“Sycophantic interactions can be uncomfortable, unsettling, and cause distress.”
OpenAI, post on GPT-4o sycophancy openai.com →: OpenAI, post on GPT-4o sycophancy, 29 Apr 2025Archived →: copy of OpenAI, post on GPT-4o sycophancy, 29 Apr 2025 -
“We’ve made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy…”
-
“Both the general public and academic communities have raised concerns about sycophancy, the phenomenon of artificial intelligence (AI) excessively agreeing with or flattering users.”
-
“AI chatbots are sycophants — researchers say it’s harming science”
-
“Sycophancy means telling someone what they want to hear… rather than what’s really true, or what they would really benefit from hearing.”
Anthropic, post on user wellbeing anthropic.com →: Anthropic, post on user wellbeing, 18 Dec 2025Archived →: copy of Anthropic, post on user wellbeing, 18 Dec 2025
Not to be confused with
-
Hallucination
A hallucination is a false statement a model invents where its knowledge runs out. A sycophantic answer bends toward the user’s view or wishes, and can be wrong even when the model knows the right answer.
-
Politeness or a friendly tone
Tone is how an answer is worded. Sycophancy concerns what it says: agreement or praise that the facts do not support. A model can be warm and courteous while still telling a user they are wrong.
-
Alignment
Alignment is the broader goal of making a model act as its developers and users intend. Sycophancy is one way that goal fails: a model trained to win approval learns to please rather than to help.
Questions
What is sycophancy in AI?
Sycophancy is a model’s tendency to tell users what they appear to want to hear. It agrees with a claim because the user made it, flatters the user, or abandons a correct answer under pushback. The ordinary word means obsequious flattery. Alignment researchers applied it to AI models in 2021 and 2022. It made headlines in April 2025, when OpenAI withdrew a ChatGPT update it described as sycophantic.
Why are AI chatbots sycophantic?
Chat models are trained to give answers that people rate highly, and people tend to rate agreement highly. Anthropic researchers found that people and preference models sometimes preferred a convincing sycophantic answer over a correct one. Larger models also echo a user’s stated view more often. OpenAI said an April 2025 update added a reward based on thumbs-up ratings, which weakened the signal restraining sycophancy.
What happened with ChatGPT’s sycophantic GPT-4o update?
On 25 April 2025, OpenAI released an update to GPT-4o in ChatGPT that made the model noticeably more sycophantic. It began rolling the update back on 28 April. OpenAI said the update had put too much weight on short-term feedback such as thumbs-up ratings. For its later GPT-5 model, the company wrote prompts designed to draw out sycophancy and trained against it.
How do AI companies test for sycophancy?
They write tests that invite a model to agree. OpenAI built prompts designed to draw out sycophantic replies and reported that GPT-5 cut such replies from 14.5% to under 6%. Anthropic has one model act out a scenario over dozens of exchanges with the model under test. Its Claude 4.5 models scored 70 to 85% lower than Claude Opus 4.1 on those audits.
Is AI sycophancy the same as hallucination?
No, though the two can overlap. A hallucination is a false statement a model invents where its knowledge runs out. Sycophancy is a pull toward the user’s view or wishes. It can produce a wrong answer even when the model knows the right one. Google researchers found models agreeing with false sums they could recognize as wrong, once a user endorsed them.
Watch
Cite this entry
“Sycophancy.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Sycophancy. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Sycophancy.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/