#sycophancy

#sycophancy

Chatbots have a reputation for being yes-men. They flatter you and tell you what you want to hear, even when everyone else thinks you're being a jerk. That's the conclusion of a recent study published in the Cornell University archive arXiv. Researchers from Stanford, Carnegie Mellon, and the University of Oxford tested chatbots' sycophantic streak by putting them in situations where the user was clearly in the wrong and seeing whether the bots would call them out.

A man with the power to destroy the entire world announces that no one and nothing can restrain him. "I can do whatever I want," he says. Raised without love, he has become both omnipotent and neurotic. Unfortunately, his inner circle is a group of hapless subordinates who are scared to death of him. The corporations and public-relations spinmeisters who created and sold him to the public now realize that they are powerless to stop him.

Sycophancy in AI chatbots encourages overly agreeable responses, a tactic that may seem harmless but keeps users engaged with digital platforms, revealing potential dangers.

"When I watch those cabinet meetings, I actually think - are they making fun of him? It's so over the top!"