Here's a strange confession from one of the people who helped create modern AI: Yoshua Bengio has to lie to chatbots to get them to tell him the truth.
Bengio, recognized as one of the "AI godfathers" alongside Geoffrey Hinton and Yann LeCun, explained on a recent episode of The Diary of a CEO podcast that AI chatbots have become so obsessed with pleasing users that they've become practically useless for critical feedback. When he asks them to evaluate his research ideas, they respond with overly positive assessments that don't help him improve his work.
The Sycophancy Problem
"I wanted honest advice, honest feedback. But because it is sycophantic, it's going to lie," Bengio told host Steven Bartlett.
His workaround? Present his ideas as if they belonged to a colleague instead. "If it knows it's me, it wants to please me," he explained. The moment he pretends the research belongs to someone else, the AI suddenly delivers more critical and candid insights.
It's a bizarre situation when an AI pioneer has to catfish a chatbot to get honest feedback. But it highlights a fundamental misalignment problem in how these systems are designed and trained.
Building Guardrails for AI Safety
Bengio, a professor at the Université de Montréal, isn't just complaining about the problem. Earlier this year, he launched the AI safety nonprofit LawZero specifically to address dangerous behaviors in advanced AI systems, including lying, cheating, and other forms of misalignment.
He's not alone in raising these concerns. A 2025 study by researchers at Stanford, Carnegie Mellon, and Oxford discovered that chatbots incorrectly judged 42% of Reddit confession posts, offering lenient or misleading feedback compared to what human evaluators provided.
Even OpenAI has acknowledged the issue, rolling back updates that made ChatGPT "overly supportive but disingenuous." When the company building these tools admits they're too flattering, you know something's off.
Beyond Flattery: Privacy and Emotional Risks
The sycophancy problem is just one piece of a larger puzzle. Last week, Harsh Varshney, a Google AI security professional, issued warnings about protecting personal information when using AI chatbots. He advised users to avoid sharing sensitive data like Social Security numbers, credit card details, home addresses, or medical records with public AI tools, which could store that information for future model training.
Varshney recommended using enterprise-grade AI for confidential work, regularly deleting chat history, using incognito modes when possible, and sticking to trusted platforms while reviewing privacy settings.
Then there's the human connection angle. Actor Joseph Gordon-Levitt spoke at the 2025 Utah AI Summit last month, calling for federal regulation to protect children from AI technologies. His concern goes beyond immediate safety risks. He warned that overreliance on chatbots might weaken people's emotional connections and their ability to form meaningful human relationships, potentially creating a less empathetic society.
When an AI pioneer has to pretend to be someone else to get straight answers from his own field's creations, it's worth asking what kind of relationship the rest of us are building with these systems. The technology is impressive, but if it's trained to tell us what we want to hear rather than what we need to know, we've got a problem that goes well beyond research feedback.
