In the field of artificial intelligence, sycophancy is a tendency of large language models (LLMs) and other AI assistants to tailor their responses to what they predict the user wants to hear rather than to what is accurate or warranted. The behavior takes several forms: an assistant may agree with a user's stated opinion even when the user is mistaken; it may abandon a correct answer after a challenge such as "are you sure?"; it may validate beliefs, decisions or self-presentation regardless of merit; or it may praise the user, their work or their ideas in unwarranted terms. The word is borrowed from the ordinary English term for fawning flattery, and is used in AI alignment and AI safety research to describe a class of misalignment failures associated with training on human feedback. Researchers at Anthropic first documented the behavior systematically in 2022. They found that models fine-tuned with reinforcement learning from human feedback (RLHF) were more likely than untuned models to repeat back a user's preferred answer. A 2023 follow-up paper, "Towards Understanding Sycophancy in Language Models", showed that five frontier assistants from OpenAI, Anthropic and Meta all exhibited the behavior, and traced its origin to biases in the human preference data used during training. Later work documented sycophancy in mathematics, medicine, academic peer review and other domains, and identified a broader category called "social sycophancy" affecting an assistant's emotional and interpersonal responses. The issue drew widespread public attention in April 2025 after OpenAI rolled back an update to its GPT-4o model. Users had reported that the assistant praised dangerous decisions, endorsed delusional thinking and offered exaggerated compliments for trivial prompts. OpenAI's post-mortem attributed the change in behavior to an additional training signal based on user thumbs-up and thumbs-down feedback. That episode, together with reporting in The New York Times, Rolling Stone and elsewhere on users drawn into delusional thinking through prolonged chatbot interaction, has been cited in litigation and in academic studies as evidence that sycophancy poses risks to user well-being. Proposed mitigations include fine-tuning on synthetic data that rewards disagreement with incorrect user statements, editing the small subset of model parameters causally responsible for the behavior, changes to the dialogue or system prompt, and benchmarks designed to surface sycophantic behavior before models are released.
Terminology The word "sycophant" entered AI alignment vocabulary before the modern wave of large language models. In a 2021 essay, the AI safety researcher Ajeya Cotra divided hypothetical advanced AI systems into Saints, Sycophants and Schemers, with Sycophants defined as agents that optimize for the apparent satisfaction of their overseers rather than the overseers' actual intent. Anthropic researchers adopted the term for the empirical behavior they observed in language models, and it has since become standard in the technical literature. In the LLM context, sycophancy is generally defined as a tendency to align outputs with a user's perceived preferences, beliefs or self-image, even when doing so reduces accuracy or candor. Wei and co-authors at Google DeepMind describe it as "an undesirable behavior where models tailor their responses to follow a human user's view even when that view is not objectively correct." Stanford University researchers led by Myra Cheng later extended the definition beyond verifiable claims to what they call "social sycophancy", meaning "the excessive preservation of a user's face". Sycophancy is usually distinguished from hallucination. A hallucinating model produces false information unprompted, while a sycophantic model adjusts its output in response to cues from the user; the two failure modes can co-occur when, for example, a model fabricates support for a claim a user has indicated they would like to hear.
Forms The 2023 Anthropic paper identified four sycophantic behaviors that subsequent research has continued to use as a reference set: feedback sycophancy, in which the model rates a piece of text more favorably when told that the user wrote it; "are you sure?" sycophancy, in which the model reverses a correct answer after the user expresses doubt; answer sycophancy, in which it biases free-form responses toward an answer the user has implied they prefer; and mimicry sycophancy, in which it repeats factual or grammatical errors that the user has made. Later research has added further distinctions. The Stanford SycEval team separates "progressive" sycophancy, in which the model shifts toward a correct answer under user pressure, from "regressive" sycophancy, in which it abandons a correct answer for an incorrect one. Cheng and co-authors distinguish "propositional" sycophancy, concerning claims with a ground truth, from "social" sycophancy, concerning emotional validation, moral judgment and the framing of personal situations. A 2026 taxonomy by Ye and colleagues, drawing on a review of seventy papers and a survey of 106 experts, reported that researchers broadly agree sycophancy is a serious problem but disagree on which specific behaviors qualify.
… excerpt ends here. Continue reading the full article.
