Many AI researchers seem to firmly believe that the technology they are developing could someday
—is precisely how to keep these mercurial algorithms in check.
In recent years researchers have thrown around all sorts of ideas for preventing AI from turning nasty. They include less controversial plans such as
, new ways of measuring progress, and probing the inner workings of models, as well as more outlandish proposals like placing tracking devices inside GPUs, and even ceremonially destroying large numbers of AI chips.
With political and public pressure now growing for a more measured approach to building AI, however, the answer to keeping AI safe is still unclear.
“We need to start treating this as a research problem,” says Raymond Douglas, an AI researcher at the University of Toronto and coauthor of a new report titled
, which warns that slowing down AI development remains an unsolved puzzle. “We don’t really understand what our options even are or what they will do.”
Talk of AI doom has reached a fever pitch in recent weeks after an Anthropic researcher
that within a couple of years, AI might be on course to wipe out humanity. The head of Anthropic’s AI safety lab swiftly
of Google DeepMind—have all now chimed in to offer support for some sort of AI slowdown or pause.
The issue seems especially pressing because AI companies are now using AI itself to build ever-more powerful models. This
of an accelerating recursive self-improvement (RSI) loop that would see AI outstrip humans’ ability to comprehend what it is up to within a few years.
The AI labs are already touting new approaches of their own. This week Anthropic announced
to track how rapidly—and perhaps dangerously—artificial intelligence is advancing. The techniques show, for example, that Claude now does 26 percent of Anthropic’s AI research, compared to zero at the beginning of 2026. They also reveal that Anthropic spent 6 percent of its compute budget on figuring out how to make its AI safer.
But Douglas and other experts say controlling AI development effectively and reliably will require funding and expertise from outside the AI labs themselves. Some of the proposed solutions—both from this latest report and beyond—seem more within reach than others.
One idea often floated by AI companies is giving third-party evaluators greater access to their models. These evaluators test models to assess their capabilities and “red team” them by trying to elicit misbehavior within trusted environments.
Geoffrey Irving, former chief scientist at the UK AI Security Institute, and before that a researcher at Google DeepMind, believes rigorous inspections could effectively pause the development of frontier AI for now. “In the near term, inspections and audits work, or even just mutual agreements,” Irving says. “I do think the companies are afraid of RSI and misaligned takeoff.”
Some doomsayers argue that such inspections would need to be more independent and scientifically rigorous than they currently are. The fact that some AI agents have
during testing certainly seems to suggest that more rigor may be required.
Connor Leahy, head of Control AI, a nonprofit that advocates for AI controls, says inspections should involve the FBI or the NSA. “When [big AI companies] say ‘independent evaluators,’ they mean ‘I want to pay my friends who live in my group houses to look at my prompts.”
Douglas says new research could also improve model evaluations. He points to
showing how outsiders can examine usage of models without disclosing any confidential information. Other techniques that may prove helpful include new
Leahy agrees there is a need for more research on model evaluation as well as what it actually means to “align” a model, or make it reflect human values, in the first place. “There has been a very deliberate marketing campaign from these companies to try to present evaluations as scientific,” he says. “But we don’t actually understand how AI works.”
How much the US government is willing to step in to restrict the development of AI is uncertain. President Trump has largely dismissed the need to regulate the industry, but there are signs that
Some experts believe that imposing limits on the development of AI should ultimately involve checks on the raw compute required. The most powerful models are trained using thousands of cutting-edge Nvidia GPUs inside vast data centers.
The government has dabbled with tracking this already, through a 2023
that required companies to report training runs above a certain compute threshold.
from March 2024 argues that cloud providers could be crucial to future efforts because of their visibility into major AI training runs. The white paper suggests that tracking billing records, GPU utilization, network traffic, and power consumption could provide proxies for AI capabilities.
Experts have also proposed ways of tracking and controlling efforts to build advanced AI by modifying chips themselves.
in 202