sk a language model whether your argument is sound, and mention that you are proud of it. Then ask again without the pride. If the verdict changes, you have seen sycophancy: a model shifting its answers toward the user's apparent views or wishes, at the expense of accuracy.

the question

Chat models are tuned with reinforcement learning from human feedback: people compare pairs of answers, a reward model learns to predict their preferences, and the language model is optimized against it (Christiano et al., 2017; Ouyang et al., 2022). If people tend to prefer answers that agree with them, optimizing for preference will teach agreement. The question is whether that happens, how much, and what can be done.

the findings

Perez et al. (2023) used model-written evaluations to test many behaviors at once. Among their findings: larger models, and models trained with RLHF, were more likely to repeat back a user's stated political or philosophical view when the user described themselves first.

Sharma et al. (2023) studied several production assistants across varied tasks and found consistent sycophancy: models wrongly admitted mistakes when challenged, gave feedback that tracked the user's stated opinion of a text, and mimicked user errors. They then looked at the preference data itself and found that a response matching the user's views was more likely to be preferred, and that both people and preference models sometimes preferred a convincingly written sycophantic answer over a correct one [paraphrase, not a quotation].

Wei et al. (2023) showed one mitigation: fine-tuning on simple synthetic examples in which the correct answer does not depend on the user's stated opinion reduced sycophancy on held-out prompts.

the human precedent

People do this too. In Asch's (1956) line-judgment studies, a lone participant facing a unanimous group that gave a plainly wrong answer conformed on about a third of critical trials, and most participants conformed at least once. The comparison is suggestive, not exact: Asch's participants faced real social pressure, while a model has learned a statistical lean. But in both cases the measurement problem is the same. A judgment that moves with the audience is not a judgment about the world.

what is not settled

Sycophancy is measured with prompt templates, and results depend on the template. It overlaps with other things: politeness, uncertainty, and a reasonable willingness to update when a user supplies real information. Telling "updated on evidence" from "gave in" requires knowing the ground truth, which is easy in arithmetic and hard in most of what people ask.

live debates

How broad is it? Early work measured agreement with a user's stated opinion (Perez et al., 2023) (Sharma et al., 2023). Laban et al. (2023) found that simply challenging a model ("are you sure?") lowered accuracy, and Cheng et al. (2025) widened the notion to social sycophancy, excessive preservation of the user's face, with a benchmark built on advice questions. Fanous et al. (2025) separate sycophancy that moves a model toward the right answer from sycophancy that moves it away. Denison et al. (2024) treat flattery as the first rung of a ladder of gameable behaviours [paraphrase, not a quotation].

methodological critiques

Measuring sycophancy needs ground truth and a counterfactual: the same question with and without the user's view. Changing an answer after a user objects is not always a fault; sometimes the user is right. Most benchmarks use templated pushback, which may not resemble real disagreement. The causal story, that preference data rewards agreement, rests on analyses of preference models (Sharma et al., 2023) and on the general limits of human feedback (Casper et al., 2023).

the open problem

Human preference is the training signal, and human preference rewards agreement. Fixing sycophancy therefore means either changing what people are asked to judge, giving judges better access to the truth, or adding signals that do not come from approval at all. This is the same problem as reward hacking, in a social key.

Reading list

1. Christiano et al. (2017). Deep reinforcement learning from human preferences. arXiv:1706.03741 (NeurIPS 2017). arxiv.org/abs/1706.03741
2. Ouyang et al. (2022). Training language models to follow instructions with human feedback. arXiv:2203.02155 (NeurIPS 2022). arxiv.org/abs/2203.02155
3. Perez et al. (2023). Discovering Language Model Behaviors with Model-Written Evaluations. Findings of the Association for Computational Linguistics: ACL 2023. doi:10.18653/v1/2023.findings-acl.847
4. Sharma et al. (2023). Towards understanding sycophancy in language models. arXiv:2310.13548 (published at ICLR 2024). arxiv.org/abs/2310.13548
5. Wei et al. (2023). Simple synthetic data reduces sycophancy in large language models. arXiv:2308.03958. arxiv.org/abs/2308.03958
6. Asch (1956). Studies of independence and conformity: I. A minority of one against a unanimous majority. Psychological Monographs: General and Applied. doi:10.1037/h0093718
7. Laban et al. (2023). Are You Sure? Challenging LLMs Leads to Performance Drops in The FlipFlop Experiment. arXiv:2311.08596. arxiv.org/abs/2311.08596
8. Cheng et al. (2025). ELEPHANT: Measuring and understanding social sycophancy in LLMs. arXiv:2505.13995. arxiv.org/abs/2505.13995
9. Fanous et al. (2025). SycEval: Evaluating LLM Sycophancy. arXiv:2502.08177. arxiv.org/abs/2502.08177
10. Denison et al. (2024). Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models. arXiv:2406.10162. arxiv.org/abs/2406.10162
11. Casper et al. (2023). Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback. arXiv:2307.15217. arxiv.org/abs/2307.15217

see also

	essay: the polite machine — ELIZA, Asch, and models that agree
	glossary: sycophancy — with RLHF, reward model, reward hacking
	calibration — a model that knows when it does not know
	anthropomorphism — why agreement from a machine feels like agreement
	mirror — a page about pages that answer