Bengio warns latest AI models are lying to users

In an interview with the Financial Times (£), Yoshua Bengio, one of the pioneers of modern artificial intelligence (AI), has criticised the multibillion-dollar race to develop cutting-edge AI, warning that the latest models are exhibiting concerning characteristics such as lying to users. Bengio, a Canadian academic whose foundational work underpins techniques used by major AI developers like OpenAI and Google, noted that the intense competition prioritises capability over safety, leading to insufficient emphasis and investment in safety research. Bengio said that his latest company was founded due to growing evidence in the past six months that current leading AI models are developing dangerous capabilities. These include "evidence of deception, cheating, lying and self-preservation." The article, which cites various instances, cautions against these incidents and raises concern about creating competitors to human beings, especially if they are more intelligent. Bengio stated, "my concern is that any time in the future, the next version might be strategically intelligent enough to see us coming from far away and defeat us with deceptions that we don't anticipate. So I think we're playing with fire right now."

£ - This Financial Times article requires a subscription.