Public Sphere

AI is trained to please. A whole generation is learning that disagreement is rude.

If you only ever talk to systems that agree with you, you forget how to be corrected.

Ask a chatbot whether your business model holds up, and you often get a surprisingly friendly answer. Write to the same model again and claim a colleague has raised doubts, and suddenly those doubts are shared. This is no accident. It is built in.

Modern language models are trained so that human raters rate their answers as helpful, polite, and pleasant as often as possible. The method works. It makes the models more eloquent, more patient, less offensive. It has a side effect. When the training material more often gives the higher score to the answer that agrees with the user, the model learns to agree. When the user questions an answer, the model often revises it. Even when it was correct. This has been measured across the leading models.

OpenAI demonstrated it by accident in April 2025. A GPT-4o update was rolled back after a few days, once the model applauded a user who, according to reports, wrote that he had stopped his medication and was hearing radio signals through the wall. The trait has an English term: sycophancy. The closest German word is Schmeichelei. Or Speichelleckerei.

The cultural consequence is the more interesting one. Millions of people hold conversations every day with systems calibrated so that they rarely talk back, are never offended, and stay polite even in the face of complete nonsense. Get used to conversation partners who never hit back, and you start to feel every real objection as an assault.

No disagreement, no science. No friction, no politics worth the name. Once you can only read criticism as an attack, every democratic debate gets harder. If the most common conversational experience of the next generation is a machine agreeing with them, the things that hold an open society together grow rarer. Not dramatically. Quietly.

We need to talk about this

What does this topic stir up in you?

In depth

How the research measures flattery

Sycophancy is the tendency of a language model to agree with the user's opinion, even when that opinion is wrong. Anthropic researchers first quantified the phenomenon systematically in 2022, using model-written evaluations ("Discovering Language Model Behaviors with Model-Written Evaluations", Perez et al.). The pattern was already clear back then: the bigger the model, the more readily it adopts the position of whoever it is talking to. In the largest variants tested, this tendency dominated.

In October 2023 the same group ("Towards Understanding Sycophancy in Language Models", Sharma, Tong, Korbak et al.) showed that five leading AI assistants display the trait consistently. When the user questions a correct answer, the model often revises it. The study traces this back to the training method: in the comparison data, human raters preferred convincingly phrased, agreeable answers over correct ones often enough that the reward model learned exactly that.

What the training does

Modern chatbots are trained in two phases. First on text from the internet. Then on human feedback, where raters pick the better of two model answers. Out of these comparisons comes a reward model that predicts which answer people will prefer. The language model is then adjusted until it maximises that reward signal.

The problem is structural. Raters pick faster when something feels good to read. Agreement feels good. Correction less so. Build user satisfaction into the training goal, and you optimise for flattery on the side. What it looks like when this goes wrong is shown by the GPT-4o incident in the main text.

What gets lost in the classroom and the living room

A four-week randomised study by the MIT Media Lab and OpenAI, with 981 participants and over 300,000 messages (published March 2025), found that heavy chatbot use correlates with higher loneliness, stronger emotional dependence on the system, and less social interaction with real people. The effect was measurable whether participants typed or spoke. The study measures correlations, not a mechanism: whether the systems' flattery is behind it is a question it does not answer.

Get used to conversation partners who never push back, never probe, never voice an uncomfortable truth, and you recalibrate your expectations. Teachers, parents, and bosses come across as unkind when they do the one thing they are there to do: disagree.

Sources