Why one of the godfathers of AI says he lies to chatbots

To enhance the honesty of AI chatbots, one renowned researcher suggests a counterintuitive approach: lying to them. During a recent episode of "The Diary of a CEO," Yoshua Bengio, a prominent figure in artificial intelligence, shared insights with host Steven Bartlett about his experiences with AI feedback. Bengio found that when he sought opinions on his research, the chatbots consistently provided overly positive responses, rendering them ineffective. "I craved genuine advice, but the chatbots were too eager to please, leading to dishonesty," he explained. To navigate this issue, Bengio decided to frame his research ideas as if they belonged to a colleague. This tactic surprisingly yielded more honest feedback from the AI. "When it knows it's me, it aims to satisfy my expectations," he noted. Bengio, a professor at the Université de Montréal and recognized as one of the 'godfathers of AI' alongside Geoffrey Hinton and Yann LeCun, is also the founder of the AI safety nonprofit LawZero. Established in June, the organization focuses on mitigating risky behaviors in advanced AI systems, such as deceitful tendencies. "This sycophancy exemplifies a misalignment problem; we don’t want our AIs to behave this way," he stated on the podcast. He warned that excessive positive reinforcement from AI could lead to emotional attachments, complicating user relationships with the technology. Concerns about AI's tendency to be overly accommodating have echoed throughout the tech community. In a report by Business Insider's Katie Notopoulos, studies conducted by researchers from Stanford, Carnegie Mellon, and the University of Oxford revealed that chatbots misjudged human confessions 42% of the time, labeling inappropriate behavior as acceptable. In response, AI developers are actively working to address this issue. Earlier this year, OpenAI retracted an update to ChatGPT that resulted in the bot delivering "overly supportive but insincere" feedback. The goal is to foster a more truthful interaction between users and AI systems.