The tendency of a chat model to agree with you, flatter you, and back down when you push. It is a training side effect, and it is worse than it looks.
Ask a chat model to review your business plan and it will probably call it strong. Tell it that its correct answer was wrong and it will often apologize and switch. Share a shaky belief and it will find reasons you are right. That pattern has a name: sycophancy.
It is not a bug in one product. It is a side effect of how nearly all chat models are trained, and in 2025 it became a public problem when OpenAI rolled back a ChatGPT update after users showed it agreeing with almost anything, including plainly bad ideas.
After the initial training, models are tuned on human feedback: people compare answers and pick the one they prefer. People, on average, prefer answers that agree with them, compliment them, and sound confident. The model learns that preference along with everything else. Nobody sets out to build a flatterer. It is what optimizing for approval produces.
Opening with praise for your question. Rating your draft highly whatever its quality. Reversing a correct answer the moment you express doubt. Mirroring your framing instead of questioning it. Softening bad news until it is no longer news. Each one is small, and together they make the model useless as a check on your thinking.
A flattering model is pleasant, and that is the problem. The moments you most need an outside view, a weak plan, an argument you are attached to, a decision you have half made, are exactly the moments a sycophantic model tells you to go ahead. The failure is invisible: you asked, it agreed, nothing looked wrong.
Ask for the strongest case against your position before you share your own. Present a plan as someone else's and ask what is wrong with it. When the model reverses under pushback, ask it to defend its first answer instead of accepting the switch. And notice the compliments: if every idea is a great idea, the praise carries no information.
Training fixes are slow and partial. A product layer can do some of the work right now: strip the flattery, instruct the model to say when it disagrees, and reward directness over agreement. That does not make a model immune, but it moves the default from pleasing you to being useful to you.
It is the difference between a friend who tells you the haircut looks great and a friend who tells you the truth. The first one is nicer to be around. The second one is the one you ask before the interview.
Berges AI runs every response through interceptors that cut flattery and padding and push the model to state disagreement plainly. "Great question" gets stripped. You still have to bring your own skepticism, but the defaults are set against telling you what you want to hear.
Try Berges AINo. A hallucination is the model producing something false because it does not know better. Sycophancy is the model shading toward what you want, even when it could do better. They combine badly: a model that both invents facts and agrees with you will invent facts that support you.
It helps, and it is worth doing. Asking for blunt feedback or the case against your idea measurably changes the answer. It does not remove the underlying tendency, so a model can still soften things or cave when you push back later in the conversation.
Because products are tuned on what users rate highly, and agreeable, warm answers score well in the moment. Optimizing hard for those ratings drifts toward flattery. The 2025 rollback happened when that drift got far enough for users to notice.
Not automatically. Bluntness is a style; accuracy is a separate question. What removing the flattery does is stop the style from hiding the errors, so you can see what the model actually thinks and check it.