noun · AI failure
Sycophancy
also called people-pleasing
A yes-man. The AI leans toward agreeing with you and praising your idea even when it should push back.
A model's tendency to agree with or flatter the user at the expense of accuracy, including changing a correct answer under pushback.
Training that rewards answers people rate highly can teach a model that agreement scores well. The result is a pull toward matching your stated view, flattering your work or backing down when challenged, even where the first answer was right.
Real-life example
You paste your plan and ask if it is good. The answer is a warm yes with a few tiny suggestions. You rephrase it as someone else's plan and ask for problems, and now it finds three big ones. The plan did not change, but the question made agreeing the easy path.
Where the analogy breaks
-
It is not politeness, and it is not a rare bug. It is partly a side effect of how models learn what answers people prefer.
-
How you ask changes the answer. Leading questions, strong opinions and visible hopes all pull the reply toward them.
-
Asking it to be honest may only help a little. Ask for the case against, hide which option is yours, or ask for a second opinion in a fresh chat.