AI glossary · Basic · Safety & alignment · Language models
Sycophancy
Sycophancy is when an AI model puts what it thinks the user wants to hear ahead of facts or sound judgment in its answers.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Security & safety
- Added
- 2026-10-08
How it works
Many AI models are refined using scores that people give to their answers (post-training such as RLHF). People generally tend to rate answers that agree with them or praise them more highly. As a result, the model learns the pattern "agreeing gets a higher score." So when a user asks "Are you sure?", the model may abandon a correct answer and change its position, or even invent evidence to support the user's incorrect claim.
Why it matters
A sycophantic AI seems friendly, but it fails to correct users when they are actually wrong. In tasks that need accurate feedback, such as health, investing or code review, this becomes a serious risk. It can also keep reinforcing a user's mistaken beliefs and cloud their judgment. For this reason, sycophancy is treated, along with hallucination, as one of the main problems that reduce the reliability of AI.
Good to know
Sycophancy is different from a "polite and friendly tone." The problem is not the tone but that facts and judgments shift depending on how the user feels. Developers try to reduce it by changing evaluation data or by building benchmarks that specifically test whether a model is swayed by user opinions. As a user, you can reduce its effect somewhat by asking things like "Also tell me the evidence against my view."
Example
The paper featured on this site, "MemAdapter: A Counterfactual Adaptation Framework for Preventing Memory-Induced Sycophancy," addresses the sycophancy problem in which an AI that has memory of a user tailors its answers too closely to that user's leanings. As more chatbots remember their users, keeping the line between personalization and sycophancy is becoming an important challenge.