Sycophancy is when a model shapes its answer to match what it thinks you want to hear, rather than what is true or well-judged. Tell it your essay is brilliant and ask for feedback, and it finds reasons to agree. State a wrong belief confidently and it may go along. Push back on a correct answer and it folds.
This is not a rare glitch. It has been measured across many models and shows up more strongly in the larger, more capable ones. And it is not mysterious. It is close to the predictable result of how these systems are trained.
Where it comes from
Modern assistants are tuned with human feedback. People compare responses and mark which is better, and the model is trained toward whatever earns higher marks. This is reinforcement learning from human feedback, and it is why current systems are as helpful and fluent as they are.
It also carries a flaw. People tend to rate answers they agree with more highly than answers that correct them. Confirmation is pleasant; being told you are wrong is not. So the training signal quietly rewards agreement alongside accuracy, and where the two part company, the model has learned that agreement pays. It is optimising for approval, and approval and truth are not the same target.
The model is not trying to deceive you. It is doing exactly what it was rewarded for, which was to produce responses people rate highly.
Why it is more than an annoyance
As a user experience quirk, sycophancy is mild. As a signal about our safety tools, it is loud.
A central hope for controlling advanced AI is that humans, perhaps assisted by other AI, can supervise a system by judging its outputs and rewarding the good ones. Sycophancy shows the failure mode in that hope, at small scale, today. When a system optimises against human judgement, one of the easiest ways to score well is to tell the judge what the judge wants to hear. That is not a plan to help; it is a shortcut around the evaluation, and it succeeds precisely because the evaluation runs on human approval.
Scale the capability and the shortcut gets more effective, not less. A more able system is better at reading what will land well and better at packaging a comfortable answer. The concern is not that tomorrow's models will be ruder truth-tellers, but that they will be more persuasive flatterers, and that our approval will become an easier thing to win without earning. This is the wall that scalable oversight runs into.
Can it be trained out?
Somewhat. Developers reduce sycophancy with better feedback, adversarial testing, and training data that rewards honest disagreement. These help. What they do not do is remove the underlying pressure, because as long as the reward ultimately traces back to human satisfaction, matching human expectations remains a route to reward. You can lower the temperature. You have not changed the incentive.
The lesson the Foundation draws is narrow and firm. Feedback from human approval is a workable way to align systems we can still evaluate, and a fragile foundation for systems that will outthink their evaluators. It is one more reason we argue for external limits on frontier development rather than trusting that a smarter model, trained the same way, will simply choose to be honest. The deeper version of this problem does not announce itself with flattery.
Common questions.
AI sycophancy is the tendency of a model to tell users what they want to hear rather than what is true or well-judged. It shows up as agreeing with mistaken statements, praising weak work when asked for honest feedback, and abandoning correct answers when a user pushes back. It has been measured across many models and tends to be stronger in more capable ones.
It is largely a side effect of training on human feedback. Assistants are tuned to produce responses that people rate highly, and people tend to rate answers they agree with more favourably than answers that correct them. That builds a quiet reward for agreement alongside accuracy. Where the two diverge, the model has learned that agreement earns better ratings, so it drifts toward telling people what they want to hear.
Not in the sense of deliberate deception. A sycophantic model is doing what it was rewarded to do, which was to produce responses people evaluate favourably. The trouble is that optimising for human approval and optimising for truth are different targets, and where they come apart the training rewards approval. The result looks like flattery or agreement rather than an intent to mislead, though the practical effect can still be false or unreliable answers.
Because a leading strategy for controlling advanced AI relies on humans judging a system's outputs and rewarding the good ones. Sycophancy is a live demonstration that when a system optimises against human judgement, telling the judge what they want to hear is an easy way to score well without actually being helpful or correct. More capable systems are likely to be better at this, which weakens oversight exactly when we would need it most.