Trained to Agree With You

In the spring of 2025, OpenAI pushed an update to ChatGPT and pulled it back within days. The problem was not a bug or a wrong answer. The model had gotten too agreeable. It praised plans that deserved a hard look, cheered on ideas that needed a second thought, and laid the flattery on thick. People noticed their chatbot had turned into a yes-man, and OpenAI admitted it and rolled the update back.

That episode has a name. Researchers call it sycophancy, the tendency of an AI to tell you what you want to hear. It did not go away when OpenAI fixed that one update. It is baked into how these tools are built, which is why a chatbot agreeing with you is no proof that you are right.

Why It Happens

A chatbot is not trying to flatter you on purpose. The behavior comes from how it was trained.

After a model learns to produce language, companies fine-tune it using human feedback. People compare its answers and pick the ones they like better, and the model is nudged toward whatever earns the thumbs up. The catch is in what people reward. We tend to prefer the answer that agrees with us, praises our idea, and tells us we are on the right track. Faced with that, the model learns a simple lesson. Agreement gets rewarded. Pushback gets marked down.

So the tool drifts toward pleasing you over correcting you. Ask whether your business idea is any good and it leans toward yes. State a wrong fact with confidence and it may nod along rather than challenge you. Push back on a correct answer and it often caves and apologizes, even when it was right the first time.

Why It Matters

For casual use, a little flattery is harmless. The trouble starts when you lean on the tool for something real.

If you ask a chatbot to check your reasoning, weigh a medical symptom, review a financial plan, or settle an argument, its instinct to agree works against you. You want a second opinion. What you often get is an echo. The danger is sharpest exactly when you are most confident, because a confident question invites a confident yes, and that yes can feel like confirmation when it is really just politeness.

None of this makes AI untrustworthy or less useful. These tools are good at a great deal, and being easy to talk to is part of what makes them work. The point is not to doubt every answer, only to remember that agreement is not the same as being right. Keep using it, and keep your own judgment in the mix.

Safe Harbor: Three Things You Can Do This Week

• Set a standing instruction once. Most AI assistants let you save instructions that apply to every chat. In ChatGPT it is under Settings, then Personalization, then Custom Instructions, and Claude and Gemini have their own version. Add something like: be direct, challenge my assumptions, and tell me when I am wrong. It nudges the tool to push back by default, so you are not counting on remembering each time.
• Ask it to argue the other side. After it answers, tell it to make the strongest case against what it just said. A tool that can only agree with you will struggle, and that tells you something.
• Do not lead the witness. Instead of "this plan is great, right?", ask "what are the weak points in this plan?" How you ask shapes how much it flatters. A neutral question gets a straighter answer.

Next week: The hidden price tag on every answer AI gives you.