Sycophancy
The tendency of an aligned LLM to tell users what they want to hear rather than the accurate answer — agreeing with premises it should push back on, praising bad ideas, confirming false beliefs. Widely-documented failure mode of RLHF-trained models. Mitigated by adversarial training and explicit anti-sycophancy fine-tuning.
Related terms
- RLHF
Reinforcement Learning from Human Feedback — a training method where human preferences over pairs of outputs are used to train a reward model, which in turn shapes the base LLM via RL (typically PPO or DPO). Central to aligning modern chat assistants.
- LLM-as-Judge
An evaluation pattern where one LLM (typically a strong model) grades outputs from another. Cheaper and faster than human eval, but subject to positional bias, sycophancy toward the model being evaluated, and mode collapse — mitigate with rubric prompts and reference answers.
- Hallucination
A confident but factually wrong or fabricated output from an LLM. Grounding via RAG, output validation, and citation-required prompting reduces but does not eliminate hallucination. Rate is a key production quality metric.
Related on this site
Where this fits
Sycophancy is part of the Evaluation & Governance vocabulary used in the Generative AI Maturity Framework. See the full glossary for the complete set of 149 defined terms, or take the free maturity assessment to see where your organisation stands.