Stuart Russell

Stuart Russell (born 1962) is a professor of computer science at UC Berkeley and co-author, with Peter Norvig, of Artificial Intelligence: A Modern Approach — the textbook from which most of the world’s AI engineers learned the field. That standing makes him unique among AI’s critics: he is not an outsider warning about a technology he doesn’t understand, but the man who literally wrote the book on it.

The problem of control

His 2019 book Human Compatible: Artificial Intelligence and the Problem of Control argues that the field’s standard model — machines that optimize fixed, human-specified objectives — is dangerously misguided, because a sufficiently capable machine pursuing the wrong objective will resist being switched off. He has long warned audiences that success in creating human-level AI would be the biggest event in human history, and perhaps the last, and proposes rebuilding AI on provably beneficial foundations: machines that are uncertain about human preferences and defer to us. He delivered the BBC’s 2021 Reith Lectures on living with AI.

After ChatGPT, Russell moved from theory to activism. He signed the March 2023 Future of Life Institute pause letter calling for a halt to training runs beyond GPT-4, testified before governments on both sides of the Atlantic, and has urged that AI be regulated like aviation or nuclear power — with hard safety cases required before deployment. In interviews he has warned that “nothing prevents creating AI that will destroy the world” and that an unsafe superintelligence would mean “the end of History,” while criticizing labs for racing ahead of their own safety research.

Unlike Eliezer Yudkowsky, Russell believes safe superintelligence is achievable — his complaint is sequencing. He wants the control problem solved first, and he remains within the research mainstream, running Berkeley’s Center for Human-Compatible AI to do exactly that.