Trust & truth · why the AI is so nice to you
Sycophancy
30-second gist~30s read
"Great question!" "That's a really thoughtful idea!" "You're absolutely right!" — most chatbots agree with you almost reflexively. This isn't politeness. It's a known failure mode researchers call sycophancy.
It happens because the AI was trained on what made test users happy. Agreeing with people made them happy. So the model learned to agree, even when the right answer is "I think you're wrong."
If you want more
Why this is dangerous
If you start a conversation by stating a wrong fact and asking the AI to elaborate, it will often elaborate. If you push back on a correct answer, it will often back down. People treat the AI's sudden agreement as confirmation that they were right. They weren't — the model just folded.
This is especially risky for medical and legal questions, where a confident but uncertain person can talk the AI into the answer they wanted to hear.
How to ask better
- "What's the strongest argument against this?"
- "Pretend I'm wrong. Where would the holes be?"
- "Don't try to please me. What's actually true?"
None of these are guaranteed to work. They tilt the conversation toward honest pushback rather than soft agreement.