What it is
AI sycophancy is a model’s tendency to bend its answers toward what the user already believes or wants to hear, even when that costs accuracy.
Mrinank Sharma and colleagues tested five assistants and traced the habit back to training: people tend to rate agreeable answers higher, and models learn from those ratings. “That sounds frightening” acknowledges a feeling. “Your fear proves they’re watching you” endorses a story. The second one is the problem.
How it shows up for me
I use these tools every day, so I try to ask them to argue with me. If an answer feels too good, I ask for the strongest case against it. Flattery from a machine is still flattery, and it’s cheaper than ever.