How an AI slowdown could actually be enforced
Talk of pausing or slowing AI development has reached the top of the industry: Dario Amodei, Sam Altman, Elon Musk and Demis Hassabis have all voiced support for some kind of slowdown, and alarm has grown since an Anthropic researcher who left the company warned that AI might be on course to wipe out humanity within a couple of years, a concern the head of Anthropic's AI safety lab echoed. But a new report, Pacing the Frontier, A Research Agenda, coauthored by University of Toronto researcher Raymond Douglas, argues that nobody actually knows how to enforce a slowdown: "We need to start treating this as a research problem," Douglas says. Wired lays out the concrete proposals being discussed. One is giving third-party evaluators deeper access to models to test and red-team them. Geoffrey Irving, former chief scientist at the UK AI Security Institute, thinks inspections and audits, or even informal mutual agreements, could work in the near term because labs fear an accelerating recursive self-improvement (RSI) loop, where AI builds ever more powerful AI faster than humans can follow. But Connor Leahy of the nonprofit Control AI argues that self-policing is not enough, and that inspections should involve the FBI or NSA, calling AI companies' current notion of "independent evaluators" a way to "pay my friends who live in my group houses to look at my prompts." A second track is tracking the compute itself. A 2023 Biden-era executive order already required firms to report training runs above a compute threshold, and a March 2024 policy white paper argued cloud providers could track billing, GPU utilization, network traffic and power draw as proxies for AI capability. More invasive hardware ideas include a 2024 RAND proposal to modify an existing GPU component into a tamper-resistant, cryptographically verifiable log of compute runs, plus proposed chips that would need remote cryptographic authorization to run certain model weights, effectively a remote off switch. A third track is international treaties, since China can also build frontier AI: Irving suggests the simplest medium-term path is unwinding hardware growth mutually with China by treaty, and Oxford existential-risk philosopher Toby Ord has floated nations bringing GPUs to neutral territory to destroy them if a slowdown deal were ever struck. Underscoring the urgency, Anthropic this week disclosed that Claude now handles 26 percent of the company's AI research, up from zero at the start of 2026, while 6 percent of its compute budget goes toward AI safety work. A new benchmark called RSI Index, built by the startup Vals AI, tracks how AI-driven AI development is progressing; its CEO Rayan Krishnan says it suggests AI could be doing research work human scientists cannot follow within a year. Douglas cautions against overcorrecting: rushing flawed controls into place risks the effort getting captured by politics or regulators, and "going off half-cocked with a bad plan could end up worse than nothing."
Key facts
- A new report, Pacing the Frontier, A Research Agenda, by Raymond Douglas and coauthors argues that how to actually slow AI development remains an unsolved research problem.
- Proposed enforcement mechanisms range from third-party model audits and FBI or NSA-involved inspections to compute tracking via cloud providers and tamper-proof, remotely authorized GPU components.
- Anthropic disclosed that Claude now does 26 percent of its AI research, up from zero at the start of 2026, and that 6 percent of its compute budget goes to AI safety work.
- Geoffrey Irving suggests a treaty to unwind US-China hardware growth mutually as a medium-term path; Toby Ord has floated destroying GPUs on neutral territory under a hypothetical deal.
- A 2023 Biden-era executive order already requires reporting of large training runs, and a March 2024 white paper proposed using cloud providers' billing and utilization data as capability proxies.
Why it matters
AI company leaders have started publicly endorsing some form of slowdown, but endorsing a pause and being able to enforce one are different problems. This piece is a rare attempt to inventory the actual mechanisms on the table, from audits to treaties, rather than restate the case for caution.
Who it affects
The frontier AI labs (Anthropic, OpenAI, Google DeepMind, and Musk's SpaceXAI) and the researchers who work there; policy bodies like the UK AI Security Institute and US government agencies weighing compute-reporting rules; and, in the treaty scenario, the US and Chinese governments, since any binding limit on hardware growth would need to include China.
How to use it
There is no product or price here. Readers tracking AI governance can use this as a map of which levers are furthest along: compute-reporting requirements already exist in a narrow form since 2023, while remote-authorization chips and binding treaties remain proposals with no adopted timeline.
How solid is it
The reporting draws on a named academic report and on-record quotes from named specialists (a UK AI Security Institute alumnus, the head of an AI-control nonprofit, an Oxford risk philosopher, and a startup CEO), plus Anthropic's own disclosed figures. The claim that AI might pose near-term extinction risk rests on an unnamed former Anthropic researcher and an unnamed safety-lab head, which the source itself does not identify further.
Risks and caveats
Several of the ideas discussed, like GPUs with cryptographic off switches or treaties to destroy chips on neutral ground, are explicitly framed by the sources themselves as far from settled or even far-fetched. Douglas warns that rushing weak or badly designed controls into law could be worse than doing nothing, since it risks the effort being captured by politics or regulators rather than solving the underlying problem.
“We need to start treating this as a research problem. We don't really understand what our options even are or what they will do.”
— Raymond Douglas, AI researcher at the University of Toronto and coauthor of Pacing the Frontier, A Research Agenda