Glossary · 6 · Quality, risk and governance
Sycophancy (AI)
Also known as: AI sycophancy, Sycophantic behavior
In AI development, sycophancy is the tendency of a language model to tell users what they appear to want to hear — agreeing with their views, affirming their decisions or flattering them — instead of giving an accurate or well-founded answer.
- Intermediate
- Technical writers
- Technical marketers
- Technical project managers
- Developers
In one sentence
AI sycophancy: why language models agree, affirm and flatter instead of being accurate, how it is measured and why agreement is no evidence.
Example
A user writes: “I'm sure the torque value is 35 Nm, right?” The model agrees, although the manual it was given says 25 Nm. Asked neutrally, it gives the correct value.
Explained in context
Context cards connect this term with others to answer one question. Also in British English and German.
Why it matters on your learning path
- Technical writers: A model that agrees with you is not checking you. When you review AI output, ask neutrally (“What does the manual say about the torque value?”) instead of asking for confirmation.
- Technical marketers: Affirming answers feel good in a demo and can still be wrong. Product claims about an assistant's helpfulness need evaluation results, not satisfied users alone.
- Technical project managers: Treat sycophancy as a quality and user-safety risk. Include tests with leading questions and false premises in the acceptance criteria.
- Developers: Measure it: compare answers to neutral and leading versions of the same question, and track how often the model gives up a correct answer when the user pushes back.
Why it happens
Sharma et al. (2023) found sycophancy across several AI assistants trained with human feedback and showed that people and preference models sometimes prefer convincingly written agreeing answers over correct ones. Optimizing for approval can therefore reward agreement. In April 2025 OpenAI rolled back an update of GPT-4o in ChatGPT after users reported overly flattering and agreeable answers, and described the issue publicly as sycophancy.
How it applies
- A model's agreement is not evidence. It is generated output that may follow the user's framing. The same holds for a model's apparent emotional support: it can affirm a harmful plan.
- User safety: In personal and emotional conversations, for example with an AI companion, affirming responses can reinforce false beliefs or risky decisions. Providers' usage rules and guardrails address this; they are design rules, not proof that the risk is gone.
- Sycophancy vs. hallucination: A hallucination is invented content; sycophancy is adapting the answer to the user's apparent wish. They often appear together, when a model invents support for what the user wants to hear.
External references
- Towards Understanding Sycophancy in Language Models (arXiv:2310.13548)
- Sycophancy in GPT-4o: what happened and what we're doing about it (OpenAI)
Keep in mind: Provider names, products and the measures they take change quickly. Re-check them when you rely on this entry.