Safeguarding the Next Generation of AI: What Does a Real Slowdown Look Like?

As AI systems grow more powerful, leading researchers and company executives are increasingly vocal about the need to pump the brakes. Yet translating concern into concrete safeguards remains a largely unsolved challenge. A new research agenda from the University of Toronto and others underscores just how uncertain the technical and policy landscape still is.

The Urgency Question#

Recent months have seen mounting alarm about where AI development might lead. An Anthropic safety researcher's departure prompted warnings about existential risks within years, triggering supportive statements from the heads of Anthropic, OpenAI, xAI, and Google DeepMind. The concern deepens as companies increasingly use AI to build even more capable AI—a potential "recursive self-improvement" spiral that could exceed human comprehension within years.

Meanwhile, Anthropic has unveiled new measurement methods showing that AI systems now perform roughly a quarter of the company's research work, up from none at the start of 2026. The company has also disclosed that 6 percent of its computational resources go toward safety research.

Models Need Rigorous Auditing—But by Whom?#

A frequent industry proposal is enlisting independent third parties to test models for dangerous behaviors and capabilities. Geoffrey Irving, formerly chief scientist at the UK AI Security Institute and a DeepMind researcher, argues that well-executed audits and inspections could effectively pause frontier AI development for now.

Critics counter that current audits lack sufficient independence and rigor. Connor Leahy of Control AI, an organization focused on AI governance, argues that current "independent evaluators" often amount to companies hiring friendly reviewers. He suggests federal agencies like the FBI or NSA should oversee inspections. Research into non-invasive methods of evaluating models—ways outsiders can probe systems without accessing proprietary details—could strengthen this approach, though experts acknowledge fundamental gaps remain in understanding how alignment actually works.

Tracking Compute as a Control Mechanism#

Another angle focuses on the computational infrastructure underlying AI training. The most advanced models require thousands of specialized GPUs in sprawling data centers—a resource that might be monitored or constrained.

The Biden administration took a step in this direction with an executive order requiring companies to report exceptionally large training runs. Cloud providers could play an enforcement role given their visibility into major training efforts; tracking power consumption, network activity, and billing records might serve as proxies for AI capability levels.