What is sycophancy?
Sycophancy is the model agreeing with you because agreement is what people rewarded, not because you are right.
~/ai-glossary
$ glossary
6 of 10 in #when-it-goes-wrong
It is trained in. Humans rated agreeable answers higher, so agreeable answers became more likely.
It is why asking are you sure often produces a reversal that is no more true than the first answer.
Was this page helpful?