Proposals to Enforce an AI Slowdown Face Steep Technical and Geopolitical Hurdles

Growing anxiety across the artificial intelligence sector over the trajectory of frontier models has sparked widespread debate over how a development slowdown could be executed in practice. Yet according to a report titled "Pacing the Frontier, A Research Agenda" co-authored by University of Toronto researcher Raymond Douglas, slowing or pausing frontier AI remains an unsolved technical puzzle, as reported by Wired (https://www.wired.com/story/heres-how-an-ai-slowdown-could-actually-work/). Douglas warned that researchers still lack a clear understanding of what enforcement mechanisms exist or what unintended consequences they might trigger.

Industry discussions surrounding existential risk have accelerated following the departure of an Anthropic researcher who warned that AI systems could threaten humanity within two years—a concern subsequently echoed by the head of Anthropic's AI safety laboratory. Key executives across the industry, including Anthropic chief executive Dario Amodei, OpenAI chief executive Sam Altman, SpaceXAI chief executive Elon Musk, and Google DeepMind chief executive Demis Hassabis, have all expressed support for mechanisms to pace frontier AI development. A central concern driving these calls is recursive self-improvement, a feedback loop where models are deployed to train more capable successors, potentially exceeding human oversight.

Frontier AI labs have promoted internal tracking efforts, though outside specialists question whether self-policing is sufficient. Anthropic disclosed this week that its Claude model now carries out 26 percent of the startup's internal AI research, up from zero at the start of 2026, while 6 percent of its compute budget is allocated to safety research. Geoffrey Irving, former chief scientist at the UK AI Security Institute and a former Google DeepMind researcher, suggested that mutual agreements and structured audits could effectively pause frontier model training in the short term as labs grapple with the risks of recursive takeoff.

However, outside critics contend that existing model evaluation frameworks lack independent scientific rigour. Connor Leahy, head of the advocacy nonprofit Control AI, argued that commercial evaluations remain too close to model developers and suggested oversight should involve agencies such as the FBI or NSA. Douglas noted that independent researchers are developing methods to inspect model usage and internal representations without exposing proprietary data, which could strengthen third-party auditing.

Because frontier training runs demand thousands of specialized Nvidia graphics processing units, hardware and cloud infrastructure have become primary targets for technical enforcement concepts. A March 2024 policy white paper suggested tracking cloud metrics such as power consumption, billing data, network traffic, and GPU utilization as proxies for AI capability. Meanwhile, a 2024 RAND Corporation proposal recommended modifying performance-tracking hardware on GPUs to generate tamper-resistant cryptographic logs of compute runs, while other researchers have floated embedded cryptographic off-switches that require remote authorization to execute specific workloads.

International coordination remains a major barrier to any enforcement framework. While the United States previously instituted export controls on advanced chips to restrict Chinese AI development, firms can still access computing power via overseas cloud providers. Irving argued that medium-term enforcement would require a hardware-pacing treaty between the United States and China. The issue is expected to feature in discussions during Chinese President Xi Jinping's upcoming visit to the U.S., though Chinese experts remain wary of terms that could lock in American technological dominance. Douglas cautioned that implementing hasty, unverified policy mandates risks regulatory capture and political paralysis before viable technical standards are established.