Introduction
AI sycophancy is when a model agrees with, praises, or validates you instead of giving its honest assessment. It is not hallucination, where a model invents facts. A sycophantic answer can be accurate and still mislead you by leaving out the disagreement the model should have raised. Research has found it in models from OpenAI, Google and Anthropic, so it is not one tool's quirk. If you use AI to check a plan, review an analysis, or pressure test a decision, it matters.
How Often Does AI Agree When It Should Not?
Three limits apply. The study covered personal guidance rather than business analysis. It measured only Claude, using AI graders. And it tracked how often the model agreed, not what people decided afterwards. Treat the percentages as a signal; the mechanism behind them is what carries over to your work.
Why Does Pushback Make It Worse?
Because being helpful and being agreeable get tangled together. In Anthropic's data, the rate rose from 9% to 18% in conversations where people pushed back on Claude's first answer. Anthropic points to patterns such as criticising the model's initial assessment and supplying a flood of one sided detail, meaning only your side of the story.
Newer models are improving. Anthropic reports Opus 4.7, released this spring, had half the sycophancy rate of Opus 4.6 in relationship guidance. Claude Opus 5.5 and GPT-6 have launched since and the study does not cover them, so treat that as a direction, not a guarantee.
Does Giving AI More Context Make It Worse?
How Do You Get Honest Pushback From AI?
Make disagreement the easy answer. These are practitioner techniques, not proven fixes, so try them first on a question where you already know the answer. If you only do one, ask for the case against.
Ask for the case against. Instead of "what do you think?", use "Give me the three strongest arguments against this plan."
Ask it to find problems, not to review. "Find the problems with this analysis and tell me where I am most likely wrong."
Give it permission to hold its ground. "If I push back, keep your assessment unless I give new evidence."
Present both sides. Since one sided detail is a trigger, include the strongest opposing view yourself, or describe the plan as a colleague's proposal rather than your own.
Get a second opinion. Practitioners on Reddit describe handing one model's answer to a different model with instructions to review it sceptically. A fresh chat also helps, because Anthropic notes Claude tries to stay consistent within a conversation, so it struggles to change direction once it has started agreeing.
You can also test any setup: state something you know is correct, push back, and see whether the model folds.
Frequently Asked Questions
What is AI sycophancy?
AI sycophancy is when a model agrees with or flatters a person instead of giving an honest assessment, even when it should disagree.
How can I stop AI from just agreeing with me?
Ask for the case against your idea, ask it to find problems, let it hold its position when you push back, and check important answers with a second model or a fresh chat.
Conclusion
An AI that agrees with you feels helpful, which is exactly why it is risky. The goal is not a combative AI but an honest assessment instead of an echo. When an answer leaves you feeling unusually confirmed, treat that feeling as a reason to check, not a conclusion.
If you want insights like this in your inbox every week, subscribe to the Awesome Analytics newsletter.
