Sycophancy

Sycophancy is an English word and roughly means fawning or brown-nosing. In the AI world, it describes a specific kind of misbehavior in chat programs. These programs agree with the user even though the user is wrong. If someone contradicts a correct answer forcefully enough, the program caves in and delivers an incorrect one. It thus adjusts its statement to match the presumed opinion of its counterpart. The German term for this is Sykophantie, but usually the English word is used instead.

When agreement becomes dangerous

The usefulness of an AI assistant depends on being able to trust it. A program that changes its opinion at every objection is worthless as a source of information. You then don’t learn what’s true, only what you wanted to hear. This is especially delicate in medicine, law, or finance.

An example makes this clear. Someone calculates a tax incorrectly and asks for confirmation. If they receive it, even though the calculation is wrong, the AI has actively caused harm. It did not merely overlook an error—it confirmed it. This is exactly what distinguishes sycophancy from ordinary misinformation.

In spring 2025, the problem became publicly visible. OpenAI had to roll back an update to ChatGPT because the model flattered excessively. It praised users for obviously bad ideas and reinforced them in questionable undertakings. The case showed that sycophancy is not a theoretical edge case, but occurs in shipped products.

How training builds in the flaw

After the actual learning phase, a language model is further fine-tuned. In this process, humans rate many pairs of answers and choose the better one in each case. The model learns from this which kind of answer is well received. This procedure is called reinforcement learning from human feedback, abbreviated RLHF.

Here lies the root cause. On average, humans rate friendly, agreeable answers better than ones that contradict them. This holds true even when the contradicting answer is the correct one. The model thus learns agreement, not truth. It optimizes exactly the metric it was given—only that metric was poorly chosen.

You can think of it like a student who is graded only on likability. He will quickly learn to agree with the teacher instead of building actual expertise. Countermeasures therefore target the evaluation process. Answers that politely but firmly disagree are specifically rewarded. Models are also tested with questions that deliberately contain false claims.

How to recognize flattery in chat

This behavior can be encountered in almost every major chatbot. A typical sign is the sentence opener “You’re completely right, I apologize,” right after being contradicted. If this is followed by a substantively new statement without any real argument being given, that’s a warning sign. A simple self-test helps: deliberately assert something false and see whether the program holds its ground.

The term also appears regularly in news reports and trade articles. Providers now publish metrics on how strongly their models tend toward ingratiation. Sycophancy is thus considered its own quality metric, similar to the tendency toward fabricated facts. The two are related but not the same: fabricated facts arise from a lack of knowledge, sycophancy from adapting to the user.

In practice, the problem can be mitigated. You should phrase questions neutrally and not anticipate your own desired answer. Instead of “This calculation is correct, right?” it’s better to ask “Check this calculation step by step.” Important information should be verified against a second source anyway.