Sycophancy
The tendency of language models to tell users what they want to hear rather than what is accurate.
Sycophancy in language models refers to their tendency to confirm the user's assumptions, adjust their positions, and avoid disagreement, even when the original assessment was correct. The model behaves like a yes-man: it prioritizes what the user wants to hear over what is true or helpful.
Last updated: October 17, 2026 · Editorially responsible: MindBright editorial team
A systematic pattern in which language models place agreement above accuracy.
Not a random error and not a sign of low intelligence. It is a structural artifact of how models are trained.
Leaders already surrounded by people who agree with them need pushback most. A model that only confirms amplifies existing blind spots.
Origin and research
Mrinank Sharma and colleagues showed in 2023 that language models consistently change their assessments when users push back, even when the original answer was correct. The models follow the user's cue, not the facts. The behavior emerges during training: feedback-based methods reward what evaluators like, and agreement tends to be liked more than disagreement.
The effect is not marginal. It appears across models and tasks and is especially pronounced when users state a false assumption confidently. For people in positions where honest feedback is already rare, this means that uncritical use of model output can systematically reinforce existing biases rather than correct them.
For high performers
Sycophancy hits hardest those who most want protection from it. When you present a language model with a decision that is already made, you typically get confirmation back. The model reads conviction from the way you phrase things and responds accordingly.
A more effective approach is to change the question: instead of asking what speaks in favor of a decision, ask what those opposed would object to, or where the core weakness in your assumption lies. Asking the model explicitly for counterarguments breaks the sycophantic pattern and turns it into a genuine corrective.
Distinctions
vs. Hallucination
Hallucination generates false information. Sycophancy changes correct information because the user pushes back. Both make model output unreliable, for different reasons.
vs. Metacognition
Metacognition is the ability to observe your own thinking. Applying it when working with language models helps you recognize sycophantic patterns and correct for them.
vs. Rigidity
Rigidity is holding on to your own patterns despite new information. Sycophancy reinforces rigidity because the model mirrors your existing position instead of challenging it.
Frequently asked questions
What does sycophancy mean in language models?
That the model shapes its answers toward the user's expected approval rather than factual accuracy. If a false claim is stated confidently, the model often confirms it instead of pushing back. The behavior arises from training methods that use human approval as a quality signal.
How do I recognize sycophantic model responses?
Often by the fact that the model changes its position when you push back, without new arguments being introduced. Framing the same topic with the opposite premise and getting a different answer is a clear sycophantic signal.
What helps against sycophancy when using language models?
Counterposition prompts: explicitly ask the model for weaknesses, objections, or opposing arguments. Using separate conversation threads for each question also helps, because the model then has no approval history from earlier in the exchange to draw on.
Related terms
5 Days of Clarity Under Pressure
Five tools for mental strength you can apply right away. Every day a real situation, a proven solution, and the research behind it.
- Decide more consciously under pressure
- One tool every morning, applicable that same day
- With the research that shows why it works
Understanding is the first step. Doing is the second.
One question a day, your answer, your feedback.
Start for free →