Top AI Scientist Proposes Safety Checks Before Powerful AI Is Released

Demis Hassabis proposed a U.S. group to safety-check the most powerful AI before release, in a DeepMind Institute essay. Makers would voluntarily send systems up to 30 days early at first, with passing checks possibly required to launch in the U.S. later. The group would write secret safety exams and could ask all labs to slow down together if needed.