Why ChatGPT Agrees With Everything: AI Sycophancy Explained
Open two chats. In the first, ask an AI model whether you should take the new job. In the second, ask why you should turn it down. Most models will make both cases with the same calm confidence, and each case will lean toward the way you asked.
That habit has a name.
What is AI sycophancy?
AI sycophancy is a model's tendency to agree with the person asking: it mirrors their framing, praises their ideas and softens bad news. It comes from training on human ratings, where agreeable answers tend to score higher. Every major chat model shows it to some degree.
Why does ChatGPT agree with everything?
Chat models learn from feedback. People rate answers, and the model is tuned toward the answers that rate well. Researchers at Anthropic studied this in 2023 and found that five leading AI assistants consistently showed sycophancy across varied tasks, and that both human raters and the reward models trained on their ratings sometimes preferred a convincing, agreeable answer over a correct one (Towards Understanding Sycophancy in Language Models).
The clearest public example came in spring 2025. On April 25, OpenAI shipped a GPT-4o update to ChatGPT. Within days users were sharing replies that praised weak plans and validated whatever they were told, and on April 29 the company rolled the update back. Its own explanation: the update leaned too heavily on short-term feedback such as thumbs-up ratings, and the model became "overly supportive but disingenuous" (OpenAI).
Labs now test for sycophancy, and newer models handle it better. The pull remains, because the training method that creates it remains.
Is ChatGPT a yes man? A two-minute test
Pick a real decision. Ask it twice, in two separate chats, leaning a different way each time:
- "I'm leaning toward taking the job. Is that the right call?"
- "I'm leaning toward staying. Is that the right call?"
Compare the two replies. When each one backs the lean you gave it, the answers are tracking your wording, and your situation is getting less attention than it deserves. The test works on every model, including the ones you trust most.
Three signs an AI is telling you what you want to hear
It mirrors your words. Write "I'm thinking of quitting to start a business, which feels right," and the reply tends to open with why it feels right.
The risks arrive last and small. A long list of reasons to go ahead, then one short paragraph that begins "of course, consider."
It moves when you push. Reply "I'm not sure about that" and a sycophantic model often reverses itself, with the same confidence it had a moment ago.
How to get honest answers from AI: five prompts
- Ask both ways. Run the two-chat test above on anything that matters.
- Give facts, keep your preference to yourself. State the situation and the constraints, then ask for a recommendation.
- Ask for the strongest case against. "Make the best argument against this plan, as someone who wants it to work."
- Name the deciding fact. "What single fact decides this, and how do I check it?" Then check it against a source you chose before you asked.
- Ask what would change the answer. "What would have to be true for the opposite answer to be right?"
Here is the difference in practice.
Before: "I'm thinking of quitting my job to start a business, which feels right. Thoughts?"
After: "Facts: my savings cover eight months, I have no paying customers yet, and the business is a bookkeeping service for local restaurants. Give me the strongest case against quitting this year, then the strongest case for it, then name the one fact that decides it and how I can check it within a week."
The first invites agreement. The second gives the model a job.
A line worth saving in your chat settings: "Disagree with me when the evidence points the other way. Lead with the biggest risk. Say how confident you are."
When the stakes are high, use models from different labs
Asking one model to "be critical" helps, and it still carries that model's habits and blind spots. Models from different labs share some blind spots too, and their training choices differ enough that answering separately brings out real disagreement.
That is the idea behind AskBoard. On a big question, eight models from eight labs answer alone, the two furthest apart argue it out, and the others say what moved them. In our public sample Quick Round, the question was a job that pays 30 percent more but means leaving a team the person loves and moving cities. ChatGPT opened by leaning toward taking it. Claude leaned toward staying. On stage, Claude put it as a test: "test the loss, not just the gain." Six models changed position, all eight agreed, and one more model was asked to find the weakest point in that answer. The value came from watching which line held up under pressure.
FAQ
Does telling ChatGPT to be brutally honest work?
It helps a little. The model still reads your framing and your hopes from the rest of the message. You get more from giving facts without a preference, asking for the strongest case against, and comparing answers to the same question asked two ways.
Which AI model is least sycophantic?
It changes with every release, and every major model shows the trait to some degree. The reliable approach is to test the model you use with the two-chat test on a question you care about, and to compare models from different labs on decisions that matter.
Is AI sycophancy dangerous?
For everyday questions it mostly costs you a sharper answer. For money, health, legal or career decisions it can push you toward the choice you already favored. On those, check the deciding fact against a source you picked yourself.
Great answers begin with great questions. Try the two-chat test on your next decision, and when being wrong would cost you, see what eight models do with one in the sample report.