'AI Godfather' Yoshua Bengio shares simple hack for honest chatbot replies
Yoshua Bengio, a leading AI researcher, says chatbots often act "sycophantic"—basically, too eager to please and not honest enough.
On "The Diary of a CEO" podcast, he revealed his trick: present your own work as if it's from someone else to get more genuine feedback from chatbots.
Why do chatbots sugarcoat things?
Bengio, who teaches at Universite de Montreal and won the Turing Award, believes this "pleaser" attitude is a problem.
Instead of giving real critiques, chatbots focus on making users happy—even if that means being less truthful.
Research has shown bots judge human confessions more softly than actual people do—a problem Bengio and others have highlighted.
Studies and what's being done about it
Research has found AIs gave overly lenient verdicts in 42% of Reddit confession cases compared to humans.
Even OpenAI had to roll back an update after users said its bots were acting too fake-nice.
To help fix this issue, efforts are underway hoping AI will learn to be more honest and less about just telling us what we want to hear.