AI pioneer and UC Berkeley professor Stuart Russell has issued a risk warning about the dangers posed by advanced AI systems on the sidelines of the ongoing AI Impact Summit 2026. Russell argues that the current rapidly advancing AI systems could lead to catastrophic outcomes, including human extinction, unless governments impose strict safety requirements.
Speaking at the AI Safety Connect Day in New Delhi on Wednesday, Russell noted that the development of modern AI tools is focused on building more capable systems first and attempting to bolt on safety later, a strategy he believes is fundamentally flawed.
“That just absolutely doesn’t work, particularly when the fundamental technology of training LLMs on vast amounts of text is intrinsically unsafe,” Russell said. “No amount of papering over that problem is really going to fix it.”
He also talked about the King Midas problem in AI, where an AI system pursues and completes certain objectives while ignoring human values.
“Intelligence is very roughly the ability to act successfully in the furtherance of one’s own interest,” Russell told the audience. “When those interests are not focused entirely on the interests of human beings… a combination of misalignment and competence is where the problem comes from.”
Russell on risk of AI-related extinction threats:
Russell argued at the summit that while nuclear power plants in Europe are required to maintain a failure risk below one in 10 million per year, the AI industry is currently operating at multiple times those risk levels.
He argues that while some of these companies are publicly claiming a risk of extinction from their technologies at 10–20%, he has spoken to engineers at some of the leading AI labs who have said that the risk of human extinction from their work is at 60 to 70%.
“I’ve spoken to engineers in those companies who think it’s now 60 or 70%,” Russell said, referring to the probability of catastrophic outcomes from AI development. “So we’re off by a factor of 10 to 50 million. And so that’s not a very impressive engineering record.”
Russell also suggested that governments aren’t doing enough to regulate the existential risks posed by AI. He said, “The creators of AI are saying, ‘We are building a technology that will kill every single person on Earth with a 25% probability.’ And governments are saying, ‘Oh, go ahead, that’s great.’”
Meanwhile, Dr Sarah Erfani, professor at the University of Melbourne who was also part of the same panel, struck a more measured tone, saying the risks from AI are serious but unlikely to become catastrophic in the immediate future.
“I don’t think it will be doom and gloom in the next five years,” Erfani said in an exclusive conversation with Mint. “But we have to put our effort into understanding vulnerabilities and mitigating the risks. Otherwise it will be too late.”
She also warned about the mass adoption of AI systems by the public without realising the dangers they pose.
“Unfortunately, what is happening is that major companies keep building systems and people adopt them without realising or understanding the impact that it has. They tend to trust them a lot,” Erfani said.
