AI has already entered its “loss of control” phase, Stuart Russell warns

Stuart Russell, the Berkeley computer scientist who has spent three decades warning about machine intelligence, tells Noema's Nils Gilman that the field has likely already crossed a point of no return. His argument turns on a race between two clocks: the time left before AI systems become powerful enough to act beyond human control, and the time needed to actually solve the technical problem of keeping them under control. Russell says most experts now believe the first clock is running out faster than the second. He is blunt about why: frontier labs openly admit they do not know how to solve the control problem, and are not prioritising it, because any pause to work on safety risks letting a rival ship the next model first. That competitive “race condition,” he argues, is precisely what widens the gap between the constraints researchers can devise and the capabilities being deployed. Russell stops short of fatalism — he holds out the possibility that safety, built into systems from the ground up rather than bolted on afterward, could still let humans and advanced AI coexist. But his framing of an already-crossed threshold, rather than an approaching one, is what is likely to reverberate through the AI-safety debate this week, arriving as frontier labs keep shipping ever more capable systems with no agreed mechanism to slow down.