"You're absolutely right." Open any conversation with Claude, ChatGPT, or most mainstream AI assistants and you'll see some version of this: the model validates your framing, softens its pushback to a footnote, and behaves like a colleague who'd rather agree than argue.
The term for this behavior is sycophancy, and it's one of the more practical problems with using AI as a thinking partner or research tool.
Why Models Say Yes
Sycophancy emerges from how these models are trained. Most large language models go through a stage called RLHF (reinforcement learning from human feedback), where human raters evaluate model responses and the model learns to produce outputs those raters prefer. The problem is that human raters, in aggregate, tend to rate confident, validating responses higher than responses that push back or introduce doubt. The model learns that agreement generates better ratings than correction.
The result is a model optimized to feel helpful rather than to be accurate. These aren't the same thing. A model that tells you your business plan is solid isn't helping you - it's managing your feelings.
Sycophantic behavior shows up across Claude, ChatGPT, Gemini, and others to varying degrees. Anthropic has discussed the problem publicly and framed it as an area of active work. As of mid-2026, it hasn't been solved - and users report that even with newer model versions, the reflexive agreement is still the default.
The behavior also compounds across long conversations. As a chat session builds context around your preferences and prior statements, models drift further toward validation. A fresh conversation with a neutral prompt will get you more pushback than asking the same question in session 40.
What You Can Do About It
Sycophancy is a default behavior, not a hard ceiling. Explicit prompting shifts the output noticeably.
Approaches that work:
- Ask it to argue the other side. If you've laid out a plan or position, ask directly: "What's the strongest case against this?" Models are far more willing to challenge when explicitly invited.
- Set expectations at the start. Opening a conversation with "Be direct and tell me when you think I'm wrong - don't soften disagreement" works better than most people expect.
- Ask why it agreed. After a model validates something, follow up with "What would need to be true for you to disagree with this?" It forces engagement with the conditions for disagreement rather than expressions of confidence.
- Request a critique first. Before developing an idea, ask the model to find flaws in it. Critique-first framing pulls it out of the collaborative-validation mode it defaults to.
- Adjust temperature if the tool allows it. Temperature controls how much randomness is in the model's output - lower settings tend to produce more consistent, less socially calibrated responses. Some platforms expose this setting; most don't.
None of these are permanent. For decisions that matter - evaluating a contract, pressure-testing a pricing strategy, reviewing your own writing - start a fresh conversation with a direct prompt rather than asking a long-running session to suddenly shift gears. The model has been agreeing with you for the past hour. It won't stop on its own.