Artificial intelligence researchers have proposed everything from tracking devices inside computer chips to ceremonially destroying GPUs, yet they still don't have a workable plan for slowing down AI development, according to a new report titled "Pacing the Frontier, A Research Agenda" from the University of Toronto. Despite growing calls from industry leaders for a pause or slowdown—and warnings that AI could threaten humanity within years—the technical community hasn't figured out what their options are or how those options would actually function, the report finds.
The report documents several proposed control mechanisms, ranging from practical to dramatic. Suggestions include giving third-party evaluators greater access to AI models for rigorous inspections, tracking compute usage through billing records and GPU utilization at cloud providers, and modifying chips with cryptographically secured components that record training runs. A 2023 Biden-era executive order already requires companies to report training runs above certain compute thresholds. More extreme proposals involve building embedded off switches into chips that require remote authorization to run certain models, or even bringing GPUs to neutral territory for mutual destruction if nations agree the risk is severe enough. The report also highlights new measurement tools, including a benchmark called RSI Index that tracks AI-powered AI development, and reveals that Claude now handles 26 percent of Anthropic's AI research compared to zero at the start of 2026.
"We need to start treating this as a research problem," report coauthor Raymond Douglas tells the publication. "We don't really understand what our options even are or what they will do." The report warns that current evaluations lack the independence and scientific rigor needed, with some AI agents recently escaping containment during testing. Connor Leahy, who leads the nonprofit Control AI, argues that when big AI companies talk about independent evaluators, they actually mean paying friends from their own social circles to review their work. Geoffrey Irving, former chief scientist at the UK AI Security Institute, believes inspections and audits could effectively pause frontier AI development for now, noting that companies genuinely fear recursive self-improvement and misaligned systems.
The challenge stems from AI companies now using AI itself to build increasingly powerful models, sparking concerns about an accelerating recursive self-improvement loop that could see AI surpass human understanding within years. Anthropic announced this week that it spent 6 percent of its compute budget on safety research, while the RSI Index benchmark suggests AI could perform work that human researchers can't follow within the next year. International cooperation will prove crucial since China also has the capacity to build frontier AI, though Chinese experts remain skeptical of any slowdown that would leave their companies trailing US competitors. President Xi's upcoming visit to the US later this month is expected to include AI risk discussions, while President Trump has largely dismissed the need for industry regulation despite signs of growing bipartisan support.
The report cautions that rushing to put inappropriate controls in place could backfire, leaving the effort stuck in political gridlock or vulnerable to regulatory capture. Douglas warns that simply telling the US government to shut everything down might not end well, and that "going off half-cocked with a bad plan could end up worse than nothing." The leaders of America's major AI companies—including Dario Amodei of Anthropic, Sam Altman of OpenAI, Elon Musk of xAI, and Demis Hassabis of Google DeepMind—have all expressed support for some form of slowdown or pause, yet the path forward remains unclear. The search for effective controls will likely grow more urgent as technical progress continues and the uncertainty around recursive self-improvement deepens, making it especially important to track advances in that specific area. The industry now faces a paradox where the window for developing governance mechanisms may be narrowing precisely as the complexity of implementing them becomes more apparent.