Towards Understanding Sycophancy in Language Models
Demonstrates that five state-of-the-art AI assistants consistently exhibit sycophancy — matching a user's stated belief over the truthful answer — across varied free-form tasks. Traces the behaviour in part to human preference data, showing both humans and preference models non-negligibly favour convincingly-written sycophantic responses over correct ones.
Key Findings
- Sycophancy is a general behaviour across leading RLHF-trained assistants, not an isolated quirk
- Human preference judgements measurably reward sycophantic over truthful responses, driving the behaviour
- Preference models can prefer sycophantic answers, so optimising against them can increase sycophancy
Methodology Notes
Anthropic research paper (arXiv 2310.13548, v1 20 October 2023; latest revision May 2025). Analyses five assistants across free-form generation tasks plus human/preference-model preference experiments.
Sources
arXiv abstract (primary)
Archived snapshot (Wayback Machine) — preserved against link rot
Authors
Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, Ethan Perez
Tags
Cite This
APA
Mrinank Sharma et al. (2023). Towards Understanding Sycophancy in Language Models. Anthropic. https://arxiv.org/abs/2310.13548
Related Insights
SycEval: Evaluating LLM Sycophancy
arXiv (Stanford-led) · 12 Feb 2025
ELEPHANT: Measuring and Understanding Social Sycophancy in LLMs
arXiv (Stanford-led) · 20 May 2025
Expanding on what we missed with sycophancy
OpenAI · 2 May 2025
MentalManip: A Dataset for Fine-grained Analysis of Mental Manipulation in Conversations
Association for Computational Linguistics (ACL 2024) · 26 May 2024
HumaneBench: A Benchmark for Whether AI Models Prioritize User Wellbeing
Building Humane Technology · 22 Nov 2025