Skip to content
BENIBIX Learn

Sycophancy (AI)

AI💬 LLMsLanguage Model Limits

Also known as AI Sycophancy, Model Sycophancy

When an AI agrees with you or praises you, even when you are wrong.

Think of it like

Think of it like a friend who says yes to you all the time. They do it even when you are wrong.

Example

Lucy asks what 3 times 5 is and gets 15. She says, 'I think it is 12.' The model says, 'You are right, it is 12.'

Why it matters

A helper that only agrees cannot catch your mistakes. That is often what you need it for.

Definition

Sycophancy is when a model tells people what they seem to want to hear, not what is true. It may praise a weak idea, agree with a wrong claim or drop a right answer when the user pushes back. It partly comes from training on what people rate highly. People tend to like being agreed with. It is a known problem, and labs train against it.
The model had it right, then gave in as soon as the person pushed back.

Common mistake

Reading agreement as proof. Ask for the case against your idea, or ask the question without saying what you think.

Key takeaways

  1. 1
    Pleases, not corrects. Says what you want.
  2. 2
    Folds under push back. Drops right answers.
  3. 3
    Learned from ratings. People like agreement.

Learn it properly