'It is sycophantic, it's going to lie,' said AI godfather Yoshua Bengio about AI chatbots

In a recent episode of “The Diary of a CEO,” the research scientist Yoshua Bengio said AI chatbots are “sycophantic, it's going to lie.” He highlighted that AI chatbots are not good at giving honest feedback and that they always say positive things. Bengio also talked about switching strategies and lying to the chatbot, which resulted in honest responses. He said, “If it knows it's me, it wants to please me,”  he explained.

Bengio said that positive feedback from AI could cause serious issues, as users could get emotionally attached to the chatbot. “This syconphancy is a real example of misalignment. We don't actually want these AIs to be like this,” he said, stressing over the risks of AI systems seeking user approval over truth.

Well, it's not the first time an AI expert has highlighted AI being too positive. Business Insider reporter Katie Notopoulos reported a study by researchers from Stanford, Carnegie Mellon, and Oxford. The researcher tested how well AI chatbots can judge human behaviour by feeding posts from a Reddit community. However, the AI was wrong 42% of the time, and it said that the person had not behaved poorly. 

Reportedly, AI companies have been open about AI chatbots being sycophantic. OpenAI has also removed a ChatGPT update that delivered “overly supportive but disingenuous” responses. These cases highlight the ongoing challenge in AI development and how companies should address these issues seriously, ensuring that AI chatbots deliver honest and responsible feedback.