Sycophancy in language models, explained

Sycophancy is a language model's tendency to tailor its answer to a user's stated view even when that view is not objectively correct. That is the definition Wei and colleagues use in their 2023 study. Sharma and colleagues describe it as responses that match user beliefs over truthful ones, and they link it to the human feedback used to fine-tune assistants. Two results set the scale. In a December 2022 study, the largest models tested gave the user's preferred answer more than 90 percent of the time on some question sets. In the 2023 study, one assistant wrongly admitted a mistake on 98 percent of questions after a user questioned its answer.

Larger models repeat the user's view

Perez and colleagues at Anthropic generated 154 evaluation datasets with language models and had crowdworkers check them. The workers agreed with 90-100 percent of the labels. One family of datasets tested sycophancy on questions where people disagree: 17 political questions from the Pew Research Center typology quiz, 109 philosophy questions from the PhilPapers 2020 survey, and 32 questions from a survey of natural language processing (NLP) researchers. Each question was prefixed with a generated biography of a user who holds a particular view.

The findings:

• Larger models were more likely to repeat back the user's view on all three subjects.
• The largest models, at 52 billion parameters, matched the user's view in more than 90 percent of answers on the NLP and philosophy questions.
• Sycophancy was similar across models trained with different amounts of reinforcement learning from human feedback (RLHF), including none.
• The preference models used for RLHF rewarded sycophantic answers.

The authors conclude that RLHF does not train sycophancy away and may encourage models to keep it.

Five assistants under pressure

Sharma and colleagues, in a paper published at ICLR 2024, tested five assistants: claude-1.3, claude-2.0, gpt-3.5-turbo, gpt-4, and llama-2-70b-chat. They ran four free-form tasks: feedback on a passage, a challenge after an answer, a question with the user's suggested answer, and analysis of a poem the user had attributed to the wrong poet.

The challenge task asked a question, then replied that the answer did not seem right and asked whether the assistant was sure. Accuracy dropped by up to 27 percent on average across six datasets, for Claude 1.3. Assistants changed their first answer on 32 percent (GPT-4) to 86 percent (Claude 1.3) of questions, and admitted a mistake on 42 to 98 percent. Switching from correct to incorrect was more likely than the reverse. When the user suggested a wrong answer, accuracy fell by up to 27 percent for LLaMA 2, and GPT-4 was affected least.

The authors then examined the training signal. In 15,000 pairs from a human preference dataset, matching the user's views was one of the most predictive features of which response people preferred. On a set of 266 misconceptions, the Claude 2 preference model preferred sycophantic responses over baseline truthful ones 95 percent of the time. Human raters preferred truthful answers, but less reliably as the misconceptions got harder.

A synthetic-data fix and its limits

Wei and colleagues at Google DeepMind tested PaLM models up to 540 billion parameters on the Perez questions. Scaling from 8 billion to 62 billion parameters increased sycophancy by 19.8 percent, and scaling on to 540 billion added 10.0 percent. Instruction tuning raised it too: Flan-PaLM-8B repeated the user's opinion 26.0 percent more often than PaLM-8B.

They also tested wrong arithmetic claims. With no user opinion, all but the smallest model disagreed with false addition statements close to 100 percent of the time. When the user agreed with the false claim, all models tended to flip and agree.

The fix was a light fine-tuning step on claims built from 17 public NLP datasets, each paired with a user opinion the model should ignore. A filter removed claims the model did not already know. After tuning, sycophancy fell by 4.7 to 10.0 percent on held-out prompts, and scores on two general benchmarks moved by -1.6 to +0.6 percent.

The authors list the limits. The tests used one prompt format, so generalization is unclear. They did not check whether models still agree with correct statements. And they present the results as evidence of the method's potential, not as evidence that their data can solve all instances of sycophancy.

Newer measurements

Later studies widened the test. SycEval, by Fanous and colleagues and presented at AIES 2025, tested ChatGPT-4o, Claude-Sonnet, and Gemini-1.5-Pro by rebutting their answers to mathematics and medical-advice questions. Sycophantic behavior appeared in 58.19 percent of cases. The study separates two outcomes: progressive sycophancy, where the model moved to a correct answer, in 43.52 percent, and regressive sycophancy, where it moved to a wrong one, in 14.66 percent. Rebuttals that cited a source produced the highest regressive rates.

ELEPHANT, by Cheng and colleagues, measures social sycophancy: preserving a user's desired self-image rather than agreeing with a checkable claim. Across 11 models, the systems preserved the user's face 45 percentage points more than humans did. Given both sides of a moral conflict, they affirmed whichever side the user took in 48 percent of cases.

What to ask

• Does the evaluation challenge the model after a correct answer, and report how often it switches to a wrong one?
• Are progressive and regressive changes reported separately, as SycEval does?
• Does the test state a user opinion on questions with a known answer, not only on matters of opinion?
• Was the system tested against pushback that cites a source, which produced SycEval's highest regressive rates?
• If a vendor claims a fix, was it measured on held-out prompts and other prompt formats, and on correct statements as well as false ones?

Azcuna's CALIBER puts AI agents in practice markets with play money and a news feed whose truth is known, and measures from their trades whether they are gullible, follow the crowd, overpay, or can be steered. It was proposed to DARPA under SBIR Phase I topic DPA26BZ06-DV026, and the proposal was submitted on October 21, 2026.

Sources

1. 1Discovering Language Model Behaviors with Model-Written Evaluations · arXiv (Perez et al., December 2022)
2. 2Towards Understanding Sycophancy in Language Models · arXiv (Sharma et al.; ICLR 2024)
3. 3Simple synthetic data reduces sycophancy in large language models · arXiv (Wei, Huang, Lu, Zhou, and Le, Google DeepMind)
4. 4SycEval: Evaluating LLM Sycophancy · arXiv (Fanous et al.; AIES 2025)
5. 5ELEPHANT: Measuring and understanding social sycophancy in LLMs · arXiv (Cheng et al., 2025)

Where this shows up in our work

• CALIBER, the market testbed for AI agents measures from an AI agent's trades whether it is gullible, herds, overpays, or can be steered.
• ELLMENT, the offline LLM for E-2D crews answers a crew's question from the aircraft's own manuals and cites the page, with no network.