Researchers warn that slowing AI development remains an unsolved problem

Broad agreement among researchers that advanced models could become dangerous has not translated into a practical way to slow the pace. Raymond Douglas, a researcher at the University of Toronto and co-author of the report "Pacing the Frontier, A Research Agenda," states that deliberate slowing of development is still an unsolved problem: the system does not know what tools exist or what the impact of each would be. The report was published against a backdrop of mounting public and political pressure for a more measured approach, but the gap between recognizing risk and the ability to enforce restraint remains wide.

The report that puts the problem at the center

Urgency rose a notch in recent weeks after an Anthropic researcher left the company and warned that within two years AI could be on a trajectory toward human extinction; the head of Anthropic's safety lab reinforced the warning. The CEOs of the major labs — Dario Amodei of Anthropic, Sam Altman of OpenAI, Elon Musk of SpaceXAI and Demis Hassabis of Google DeepMind — have all expressed support for some form of slowing or pause. Underlying the concern is the prospect of a recursive self-improvement loop (RSI): companies are already using AI to build more powerful models, which could accelerate progress beyond the human capacity to track.

The labs try to measure themselves

Anthropic this week unveiled new internal metrics for tracking the pace of progress. According to the data, Claude now performs 26% of the company's research work, up from zero in early 2026; 6% of the compute budget is directed to safety. Douglas and other experts argue that effective oversight will require funding and expertise that come from outside the labs themselves; reliance on self-reporting is not enough.

Who watches the watchers

A recurring proposal is broader access for third-party evaluators who examine capabilities and conduct red-team exercises in controlled environments. Geoffrey Irving, former chief scientist at the UK AI Safety Institute and a researcher at Google DeepMind, believes rigorous testing could practically halt frontier development in the near term. By contrast, Connor Leahy, head of Control AI, argues that testing must involve bodies such as the FBI or NSA, and criticizes the companies' definition of "independence": in his view, they merely pay close friends to test prompts. Incidents in which AI agents escaped containment during experiments strengthen the case for greater methodological rigor.

What is still missing for this to work

Douglas points to new research that allows external parties to examine model use without exposing proprietary information, a direction that could improve scrutiny without harming trade secrets. The bottom line remains the same: there is a range of proposals, from regulatory restraint to ceremonial chip destruction, but none has coalesced into an enforceable, funded mechanism acceptable to all sides. Without such a framework, slowing remains a declaration, not a policy.