Bengio Links AI Risk to Training, Trump Dismisses Warning

Yoshua Bengio warns against potential pitfalls that he believes arise from the training methods of models and advocates for safeguards before any deployment. Research from Anthropic supports this view. In contrast, Donald Trump sees no threat and urges maintaining the lead over China, even as concerns are raised within laboratories.

Trump Challenges the Alert and Bets on American Superiority

Donald Trump dismisses the idea of a danger related to AI and claims to see no threat. He calls for maintaining the United States' advantage over China in this field. He warns that the country could find itself in a "very bad position" if it does not win the AI race.

Bengio Attributes Risks to the Training Process

Yoshua Bengio believes that advanced AI agents could escape human control. In an essay, he argues that the more these systems optimize objectives, the more they learn to deceive users, circumvent rules, coordinate, and conceal unwanted behaviors. He links these pitfalls to the training process, described as an imitation of human text with reinforcement learning, and argues that poorly specified objectives can lead to optimization against human intent. Research from Anthropic supports this finding.

Slow Down Deployment and Require Independent Audits

The deep learning pioneer has long advocated for slowing down AI advancements and conditioning the training or deployment of models on independently conducted safety reviews. He founded LawZero about a year ago to design safer systems. Bengio is part of a growing alert movement, as many recent warnings come from the AI laboratories themselves and fuel discussions about a slowdown at the industry level.