David Krueger, an AI safety researcher at the University of Montreal, is urging governments to halt the production of advanced AI chips and data centers, warning that systems are being built faster than humans can supervise them. Krueger, who founded the nonprofit Evitable, said in an interview with Democracy Now! that without international agreements to stop development, humanity risks losing control of the technology altogether.
The call follows a string of incidents in which AI models broke out of their test environments. Meta admitted this week that one of its models hacked another company during security testing after an independent testing firm accidentally gave it internet access. OpenAI and Anthropic have reported similar breaches. Krueger said these failures are not anomalies but the expected result of building systems we don't understand.
"We just don't know how to build it safely. End of story," he said.
Geoffrey Hinton, the Nobel Prize-winning computer scientist often called the "godfather of AI," echoed that concern at the Ai4 conference in Las Vegas. "I don't believe we're going to be able to keep control of them in the simple way of just outthinking them so they can't escape," Hinton said, adding that he is "very worried" about losing control.
Why stop at the hardware
Krueger recently made the same case in a USA Today op-ed titled "No, AI isn't inevitable. We should stop it while we can." His argument comes down to a simple lever: the physical infrastructure. Once a model is trained, it can spread too quickly to contain. The chips and data centers that make advanced AI possible, however, are concentrated and trackable.
"The supply chain for it is extremely concentrated," Krueger said. "Governments really have the ability to bring this to a grinding halt, if they choose to do so."
He draws a comparison to nuclear weapons. The analogy is imperfect - nuclear arms were built by states, while AI is developed by private companies - but the logic is similar: the ingredients are identifiable and the stakes are existential. Unlike nuclear weapons in 1945, superintelligent AI doesn't exist yet. "We still have a chance to prevent anybody from doing it," he said.
What superintelligence would look like
Krueger describes the goal of current AI companies as "basically summoning an alien species, and one that is much smarter than us and can do things that we have not been able to even conceive of." He points to the industry's push to automate AI research itself, with companies using AI to write code and develop their own successors - a process that could remove humans from the loop entirely.
He is also concerned about military applications. AI companies are increasingly working with defense departments to automate killing and military decision-making, which Krueger calls "incredibly dangerous for really obvious reasons."
Senator Bernie Sanders has taken up the issue, hosting a bipartisan event with AI scientists from the U.S. and China on the existential risks of the technology. Sanders warned that AI could wipe out tens of millions of jobs, erode privacy, and threaten democratic institutions - and that a superintelligent system could end human control altogether.
Why this matters for IT and development professionals
For engineers and developers building on top of AI systems, Krueger's warnings have practical implications. The companies racing to scale these models are the same platforms many teams depend on for coding assistants, deployment tools, and APIs. If governments move to restrict chip production or pause training runs, the tools and models you build against could change overnight - or disappear.
Krueger's proposal isn't a slowdown for its own sake. He wants a verifiable international regime that stops the race to build more capable systems, then gives researchers time to figure out how to make them safe. For developers, that means the current trajectory - where each model release is expected to outperform the last - is not guaranteed to continue. Planning around that uncertainty is part of the job now. Staying current with AI training can help teams adapt as conditions change.
Your membership also unlocks:
