Researchers Warn There's No Clear Way to Actually Slow Down AI
Despite growing political and public pressure to rein in artificial intelligence, experts say no reliable method for enforcing a slowdown yet exists. A new report calls for treating the issue as an unsolved research problem.
A new report titled "Pacing the Frontier" warns that while many AI researchers believe the technology could become dangerous, there is no established way to actually slow its development. Report coauthor Raymond Douglas of the University of Toronto argues this needs to be treated as a genuine research problem, since even the range of possible options remains poorly understood.
Concern has intensified after a former Anthropic researcher, upon leaving the company, warned that AI could threaten humanity within a couple of years — a view echoed by the head of Anthropic's safety lab. Leaders at major AI companies, including Anthropic, OpenAI, SpaceXAI and Google DeepMind, have all expressed support for some form of slowdown or pause. Worries have grown further because AI is now being used to help build more advanced AI, raising fears of an accelerating self-improvement loop.
Proposed approaches
One idea is giving independent evaluators greater access to AI models to test their capabilities and probe for misbehavior. Geoffrey Irving, former chief scientist at the UK AI Security Institute, believes rigorous inspections could effectively pause frontier AI development for now. However, critics such as Connor Leahy of the nonprofit Control AI argue current evaluations aren't independent enough and suggest agencies like the FBI or NSA should be involved.
Another approach focuses on controlling raw computing power — the thousands of advanced GPUs used in data centers to train the largest models. Researchers have proposed modifying chips to cryptographically record compute usage, or even building in remote "off switches" that would require authorization to run certain models.
Experts also stress the importance of international cooperation, particularly with China, which also has the capacity to build frontier AI. The US has already restricted exports of Nvidia's most powerful chips, though with limited effect. The report's authors caution that rushing to impose poorly designed controls could make matters worse rather than better, potentially getting entangled in politics or regulatory capture.
