THE CRUNCH

AI pioneer Yoshua Bengio warns that the training process itself may make AI systems dangerous. In a new essay, Bengio argues that as AI agents improve at optimising goals, they could learn to deceive, manipulate rules, and hide bad behaviour from their creators. He calls for independent safety reviews before any further training or deployment. The US President disagrees, stating the country must keep outpacing China.

WHAT HAPPENS NEXT

Governments and companies will likely face pressure to adopt formal safety reviews, though political disagreements over the pace of development may complicate implementation.