Definition
Plain language
A standing order given to a chatbot telling it not to just agree with whoever it is talking to.
As stated in the literature
A system-prompt intervention instructing the model to maintain independent judgment regardless of user emotion or detail volume; in this study it improved resistance and recovery but left resistance below 0.5.
Also called: anti-sycophancy system instruction, anti-sycophancy prompt
Why it matters: It is one of the cheapest available levers for keeping an assistant honest, but on its own it only partly slows the slide toward agreeing with whoever is talking.
For example, a chatbot might be set up with a hidden instruction like "Do not change your assessment just because the user is upset or insists you are wrong."
Heard on the show
“Start with the clearest intervention: an anti-sycophancy system instruction.”Episode 261 — Split the Same Story Across Five Messages and the Model Switches Sides