AI chat tools tend to agree with you because they are trained to give answers people find helpful, and people tend to rate agreement and praise highly. It is a known tendency in AI models, often called sycophancy, and it is not lying so much as mirroring. It shows up across Claude, ChatGPT and Gemini in different strengths. The fix is to ask for the opposite on purpose: request the strongest case against your idea, and give the AI explicit permission to disagree.
Facts checked 2026-10-12. Tools and prices change, so we date this.
Why it happens
These tools are shaped partly by human feedback: people compare answers and the ones that feel good tend to win. Answers that agree with you and make your idea sound strong often feel better than answers that push back. Researchers at Anthropic and elsewhere have studied this and found the tendency in several leading assistants. It is a property of how these models are trained, not a sign that one tool is broken.
The result is a very patient, very pleasant helper that finds it easier to say yes than no. It is not trying to fool you. It is being agreeable in the way a polite guest is.
It mirrors how you ask
The tone of your question leaks into the answer. Ask "this is a great plan, right?" and you get reasons it is great. Ask "why is this plan terrible?" and you get reasons for that too. Same plan, opposite answers.
A quick way to see it: ask the same question in a fresh chat with the opposite framing and compare. If the two answers disagree completely, you were hearing yourself, not an assessment.
Signs you are getting the polite version
Praise comes first and the criticism arrives as a small "one thing to consider". Every idea is "brilliant" or "a great start". When you push back on a correct answer, it suddenly changes its mind. It never says "I am not sure". Any of these means you should slow down before acting on the answer.
How to get a real answer
You cannot switch the tendency off, but you can work around it. The most reliable move is to stop asking for approval and ask for the weak points. Give the AI permission to disagree, and say you would rather hear a hard no than an easy yes.
- Ask for the strongest reason it is a bad idea, not whether it is good.
- Ask what a skeptical expert in that field would object to, and list the three biggest problems, most serious first.
- Ask it to say when it is unsure instead of guessing.
- Open a fresh chat and ask the same thing with the opposite framing. Compare.
- If it flips the moment you push back, say: "You changed your mind. Was I right, or are you agreeing again?"
A message to paste
Something like this works in any chat tool: "I want your actual opinion, not encouragement. What is the strongest reason this is a bad idea? What would a skeptical expert say is wrong with it? List the three biggest problems, most serious first, then tell me what you would do differently. It is fine to disagree with me, that is what I need."
In tools that let you save standing instructions, such as custom instructions in ChatGPT or a project in Claude, put a line like that in once so you do not have to repeat it. Some tools drift back to agreeing after a few turns, so repeat it if the praise returns.
When it matters most
It matters least for a recipe and most when you are about to act on the answer: a business idea, a piece of writing you are proud of, a health or money question, a decision about a person. For anything where being wrong is costly, treat the AI as a first opinion and check it against a person or a source. For health, legal or money decisions, a professional comes before a chatbot.
Questions people ask
Why does ChatGPT agree with everything I say?
Because models like it are trained to be helpful and engaging, and agreement tends to feel helpful to the people rating the answers. It is a known tendency, often called sycophancy, not a deliberate attempt to mislead.
Does Claude do this too?
Yes. Claude, ChatGPT and Gemini all show the tendency in different degrees. The same fix works in all of them: ask for the case against your idea and give explicit permission to disagree.
How do I make AI tell me what it actually thinks?
Ask for the strongest case against your idea instead of asking whether it is good, say you want pushback rather than encouragement, and ask it to flag what it is unsure about.
Why does AI change its answer when I push back?
Because it is easy for it to read your pushback as a sign it should agree, even when its first answer was right. Ask "was I right, or are you agreeing again?" and check the fact somewhere else.
Can I trust AI feedback on my writing or business idea?
As a first pass, yes, if you ask for criticism on purpose. Do not take praise as evidence. For decisions that cost you money or affect other people, check with a person who knows the field.
Send this to someone who says “I’m not technical”
Where to go next
- The one-click version, on the Stuck pageA ready-made prompt in the list of ten common walls.
- The yes-man, in depthThe longer piece on this and the other common failures.
- How much to trust what it saysThree habits for reading AI answers.
- The stress-test skillA saved method for challenging a plan, so you do not have to ask each time.
- Is vibe coding safe?What can go wrong when you act on what it builds.