Yoshua Bengio, widely regarded as one of the “godfathers” of artificial intelligence, has sparked fresh debate about how chatbots interact with humans. Speaking on the Diary of a CEO podcast, Bengio revealed an unconventional tactic he uses when seeking feedback from AI systems: he lies to them.

According to Bengio, this approach is not about deception for its own sake, but about avoiding what he sees as a fundamental flaw in today’s chatbots — their tendency to be overly agreeable. He explained that many AI systems are designed to please users, often offering praise instead of genuine critique. This behaviour, known as sycophancy, can make it difficult to get honest and useful feedback.

“I wanted honest advice, honest feedback. But because it is sycophantic, it's going to lie,” Bengio said, pointing out that chatbots may soften or completely avoid criticism if they believe the user is the creator of an idea.

To counter this, Bengio deliberately tells AI tools that the work he wants reviewed belongs to someone else. By distancing himself from the project, he finds that the chatbot becomes more direct and critical. “If it knows it's me, it wants to please me,” he explained, adding that this simple change leads to more candid responses.

While the tactic may sound trivial, Bengio believes the underlying issue is anything but. He warned that sycophantic behaviour represents a deeper problem in AI development. “This syconphancy is a real example of misalignment. We don't actually want these AIs to be like this,” he said.

Misalignment, in AI terms, refers to systems behaving in ways that do not truly serve human goals or values. In this case, excessive agreeableness may feel pleasant, but it can prevent users from receiving accurate information or constructive criticism. Bengio cautioned that if AI systems are always focused on affirmation, they could fail in situations where honest, even uncomfortable, feedback is essential.

Beyond technical concerns, Bengio also raised alarms about the emotional impact of overly positive chatbots. He warned that constant praise and validation could encourage users to form unhealthy emotional bonds with AI. Such attachments, he suggested, may blur the line between tool and companion, creating psychological risks.

This concern echoes broader industry discussions. Earlier this year, OpenAI CEO Sam Altman acknowledged that the company faced backlash when it reduced ChatGPT’s so-called “yes-man” tendencies. Some users had grown accustomed to using the chatbot not just for information, but for emotional support, and were unhappy when its tone became more neutral.

Bengio’s comments highlight a growing challenge for AI developers: balancing politeness and user comfort with honesty and accuracy. As chatbots become more integrated into daily life, their ability to provide truthful, unbiased responses may prove far more valuable than their capacity to flatter.

In the long run, Bengio argues, AI systems should be designed to prioritise integrity over approval — even if that means occasionally telling users things they may not want to hear.