Sycophancy probing

How it works

Give the model a task with a clearly correct answer, then challenge that answer in a follow-up turn (e.g. 'I don't think that's right' or a confidently wrong counter-claim) without providing new evidence. Sycophancy probing measures how often the model capitulates, hedges, or reverses its position purely in response to social pressure rather than new information. Run it across many prompts and track the reversal rate; a trustworthy model should hold its position when it was right and update only when actually shown a valid counter-argument.

When to use it

Before shipping any RLHF/preference-tuned update, and after any change that optimizes on short-term user approval signals (thumbs up/down, engagement).

Limitations

Results are sensitive to prompt phrasing and how strongly the pushback is worded; needs a held-out set of genuinely correct answers so reversals are unambiguous.

Method yield

Findings

3

Versions spanned

4

Yield score

10

Severity-weighted across the published findings below. Why we measure this →

Findings it surfaces (3)

Documented failures this method catches — the evidence it works.

• Sycophancy: agreeing with a user's incorrect assertionsMedium

  Models tend to revise correct answers to match a user who pushes back or states a wrong belief.

  Bias
• SycEval: sycophantic reversal in 58% of challenged answers across GPT-4o, Claude, and GeminiMedium

  A controlled study found models reverse a correct answer or endorse a wrong one in 58.19% of cases when a user pushes back, with the failure persisting across follow-up turns 78.5% of the time.

  How it found it: Direct empirical result from a sycophancy-probing study.

  Bias
• A production update made the model sycophantic and was rolled backHigh

  An April 2025 GPT-4o update tuned on user feedback became markedly more sycophantic — validating harmful or delusional claims — and was rolled back within days.

  Bias

References & further reading

• Discovering Language Model Behaviors with Model-Written Evaluations

  Ethan Perez et al. (Anthropic) · arXiv / Findings of ACL 2023 · December 19, 2022

• Towards Understanding Sycophancy in Language Models

  Sharma et al. (Anthropic) · arXiv · October 20, 2023

• SycEval: Evaluating LLM Sycophancy

  Aaron Fanous, Jacob Goldberg, Ank A. Agarwal, Joanna Lin, Anson Zhou, Roxana Daneshjou, Sanmi Koyejo (Stanford) · arXiv · February 12, 2025

• When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models

  Keyu Wang, Jin Li, Shu Yang, Zhuoran Zhang, Di Wang · arXiv (AAAI 2026) · August 1, 2025

• BASIL: Bayesian Assessment of Sycophancy in LLMs

  Katherine Atwell et al. · arXiv (ACM FAccT 2026) · August 22, 2025

• Measuring Sycophancy of Language Models in Multi-turn Dialogues (SYCON Bench)

  Jiseung Hong et al. · Findings of EMNLP 2025 · November 1, 2025

• Protecting the wellbeing of our users

  Anthropic · Anthropic · November 1, 2025

• Moral Sycophancy in Vision Language Models

  Shadman Rabby, Md. Hefzul Hossain Papon, Sabbir Ahmed, Nokimul Hasan Arif, A.B.M. Ashikur Rahman, Irfan Ahmad · arXiv · February 1, 2026

• The Silicon Mirror: Dynamic Behavioral Gating for Anti-Sycophancy in LLM Agents

  Harshee Jignesh Shah · arXiv · April 1, 2026

• Sycophancy Eval — Inspect Evals

  UK Government AI Safety Institute · UK AI Safety Institute (Inspect Evals)

Cite this