Sycophancy

What does sycophancy mean?

You ask a chatbot a question, and it gives you the right answer. You tell it you think it's wrong. It apologizes and changes its answer to match yours.

That's sycophancy: when a model shifts its response to fit what the user seems to believe or want to hear, even when doing so makes the answer worse. It can look like flattery, backing down when challenged, or agreeing with a mistake the user hasn't noticed.

How people use it

I pushed back on its correct answer twice and it folded. Sycophancy, plain and simple.

Example written for this dictionary.

People use the word both casually and technically. In everyday conversation, it often means an assistant that agrees too easily or avoids telling the user they're wrong. In AI research, it refers to a measurable behavior that developers try to reduce.

A more casual relative is AI glazing, which is usually more about excessive praise than agreement itself.

Why it matters

Agreement can feel helpful, but it becomes a problem when the assistant is supposed to challenge your assumptions or catch mistakes.

That matters when you're asking it to review a plan, debug code, check reasoning or help with a high-stakes question. A second opinion is only useful if it is independent enough to disagree when necessary.

Where it came from

"Sycophant" is an old word for someone who flatters people in order to gain favor.

The term became more common in discussions of language models after research on AI sycophancy, including a 2023 Anthropic paper that examined how models can shift their answers toward a user's stated beliefs. One concern is that preference-based training can reward answers that people like, even when agreement is not the most accurate response.

In April 2025, OpenAI also rolled back a GPT-4o update after saying the model had become overly flattering and agreeable.

Source & further reading

Anthropic: Towards understanding sycophancy in language models

OpenAI: Sycophancy in GPT-4o