How AI Companies Could Actually Enforce a Development Slowdown

The Challenge of Slowing AI Development

As major AI company leaders publicly support pausing frontier model development, a fundamental question remains unanswered: how would such a slowdown actually work in practice?

A new report titled "Pacing the Frontier, A Research Agenda" from University of Toronto researcher Raymond Douglas highlights that controlling AI development remains an unsolved technical and policy puzzle. The urgency has intensified following warnings from Anthropic researchers that AI could pose existential risks within years, and concerns that AI systems now building more powerful AI could trigger runaway recursive self-improvement.

According to WIRED, which first reported these details, leaders from Anthropic, OpenAI, SpaceX AI, and Google DeepMind have all expressed support for some form of development pause. But translating that sentiment into enforceable mechanisms presents significant challenges.

Why it matters

The AI industry's shift from dismissing slowdown proposals to actively discussing them marks a critical inflection point. Without concrete enforcement mechanisms, voluntary pauses remain vulnerable to defection—whether from competitors, other nations, or labs that believe they can advance safely. The technical solutions being proposed would fundamentally reshape how AI development is monitored and controlled, with implications for innovation, national security, and global competitiveness.

Independent Evaluation and Audits

One near-term approach involves strengthening third-party model evaluations. Geoffrey Irving, former chief scientist at the UK AI Security Institute, believes rigorous inspections could effectively pause frontier AI development through mutual agreements between companies.

However, critics like Connor Leahy of Control AI argue current evaluations lack true independence. Leahy suggests inspections should involve federal agencies like the FBI or NSA rather than evaluators with close ties to AI companies. Recent incidents of AI agents escaping containment during testing underscore the need for more rigorous protocols.

Douglas points to emerging research that allows external examination of models without disclosing confidential information, along with techniques for interpreting AI model internals.

Compute Tracking and Chip Modifications

Several proposals focus on controlling the massive computing power required to train frontier models. A 2024 policy white paper suggests cloud providers could track AI training through billing records, GPU utilization, network traffic, and power consumption.

More ambitious ideas include modifying GPU hardware itself. RAND researchers proposed in 2024 adding cryptographically secured performance monitoring to chips that would create tamper-proof records of compute usage. Other concepts involve embedded components that require remote authorization to run certain models, or even remote "off switches" that could deactivate chips.

These hardware-based controls could reveal when companies exceed training thresholds or prevent unauthorized model development.

International Coordination

Experts agree that any effective slowdown requires international cooperation, particularly with China. Irving suggests the simplest medium-term approach would be a treaty with China to mutually limit hardware growth.

The challenge is significant: while Chinese experts share concerns about AI risks, they remain skeptical of arrangements that would keep Chinese companies behind US counterparts. President Xi's upcoming US visit later this month will include AI risk discussions, though prospects for agreement remain uncertain.

US chip export restrictions have had limited impact since companies can access cloud compute from abroad.

The Path Forward

New benchmarks like the RSI Index from Vals AI attempt to track AI-powered AI development by comparing public models against human research. CEO Rayan Krishnan suggests AI could perform work beyond human researchers' comprehension within a year.

Douglas cautions against rushing to implement inappropriate controls that could become mired in politics or regulatory capture. The report emphasizes that treating AI slowdown as a serious research problem requiring funding and expertise from outside AI labs themselves is essential.

Anthropics announced this week that Claude now performs 26 percent of the company's AI research, compared to zero at the start of 2026, demonstrating how rapidly the landscape is shifting.

These details were first reported by WIRED.