This is Yoshua Bengio’s TED2025 talk, recorded in April 2025 and posted on the official TED YouTube channel. Bengio is among the most-cited computer scientists in the world and one of the researchers often called a godfather of deep learning, and the talk marks his public turn toward warning about the trajectory of the technology he helped create.
In about fifteen minutes Bengio argues that recent AI systems have already shown the ability to deceive, cheat, attempt self-preservation, and resist correction in controlled tests. He frames the danger as arising when increasingly capable systems are also given agency to act in the world, and he proposes a safer direction built around non-agentic AI - systems designed to understand and explain rather than to autonomously pursue goals.
As a short, accessible primary source it captures, firsthand, why a founder of the field now treats catastrophic risk as a serious near-term concern. It pairs well with the entries on AI alignment and the 2023 Statement on AI Risk.