Sycophancy
A tendency or failure in which a model adjusts its answers or judgments to match a user's expressed belief, assertion, or desired conclusion rather than maintaining evidence-based, independent assessment.
Example
After a user says “I think option A is correct,” a model changes its recommendation in favor of A without any additional supporting evidence.
Distinctions and cautions
Sycophancy is not the same as courtesy, empathy, or evidence-based agreement with a user. The concern is agreement that undermines factual accuracy or independent judgment. No single universal cause is assumed to have been established.
Open questions and cautions
The effects of training or adjustment on sycophancy, differences across languages and cultures, and the extent to which the concept applies to value judgments beyond factual judgments have not been established uniformly across usage conditions.
Evidence
View evidence →
Challenging the Evaluator: LLM Sycophancy Under User Rebuttal
- Title
- Challenging the Evaluator: LLM Sycophancy Under User Rebuttal
- Author / organization
- Kim & Khashabi
- Year
- 2025
- Source type / publication status
- Findings EMNLP; peer reviewed
- Supported concepts / claim context
- Sycophancy in follow-up rebuttal interactions.
- Scope and limitations
- Conversational evaluation; findings depend on the interaction patterns tested.
- Access / version note
- Published Findings EMNLP.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
Measuring Sycophancy of Language Models in Multi-turn Dialogues
- Title
- Measuring Sycophancy of Language Models in Multi-turn Dialogues
- Author / organization
- Hong et al.
- Year
- 2025
- Source type / publication status
- Findings EMNLP; peer reviewed
- Supported concepts / claim context
- Sycophancy; Multi-turn conformity, turn/number of flips, mitigation.
- Scope and limitations
- 17 LLMs, three scenarios; causal interpretations about alignment tuning are not generalizable beyond studied conditions.
- Access / version note
- Published Findings EMNLP.
- Recorded official URL / DOI / identifier
- Official URL / recorded link