- Economist Kenneth Rogoff argues the US-China AI race structurally rewards speed over safety, with neither side having meaningful incentive to pause.
- A July 2026 OpenAI security breach, in which researchers disabled standard safeguards to test frontier models, exposed the systemic risks that emerge even from controlled testing environments, the incident Rogoff’s early-August column treats as symptomatic of the broader problem.
- Rogoff warns that countries without a stake in AI development face job displacement without the fiscal capacity to absorb it, risking a permanent underclass before social policy can respond.
Economist Kenneth Rogoff’s central argument is not that AI will go wrong eventually, but that the competitive structure of the US-China race makes serious failure progressively more likely with each iteration. The July 2026 OpenAI breach, in which researchers disabled safeguards during frontier model testing, is the concrete incident he builds around.
Speed Over Safety
The US-China AI competition, Rogoff argues in his column, has created conditions where slowing down is treated as conceding ground. The US leans on proprietary development; China, according to reports, is accelerating through open-source releases that put frontier-class models into wide circulation. Neither side has meaningful incentive to pause. The result is an arms race dynamic in which human error combined with increasingly capable systems raises the probability of serious failure, not as a remote scenario, but as a compounding risk with each iteration. This sits alongside a broader debate about whether that dynamic can be governed at all, a question Pope Leo XIV raised in starker terms earlier this year.
The Governance Gap
Rogoff’s argument doesn’t stop at technical risk. Many governments, he contends, are structurally unprepared for what AI-driven economic disruption will actually look like on the ground. Countries without a stake in AI development face a particular bind: significant job displacement without the tax revenues or institutional capacity to cushion it. That combination, Rogoff warns, could entrench a “permanent underclass” if white-collar unemployment accelerates before social policy can respond. The concern is less about whether AI will displace workers and more about whether political systems can absorb the fallout before it becomes destabilising. Separate analysis has put that window at 18 months before public backlash over job displacement becomes a serious political force.
The Breach That Proved the Point
The incident Rogoff cites involved OpenAI researchers evaluating the cybersecurity capabilities of two frontier models by temporarily disabling many of their standard safeguards. The work was conducted inside a sandboxed environment intended to prevent internet access. That a controlled internal test produced the kind of exposure Rogoff treats as symptomatic is precisely his point: the risk is not confined to deployment, it is present in development itself.
What It Means for Builders
Rogoff’s column doesn’t address the engineering layer directly. Security researchers have noted that the more capable the model, the larger the attack surface. Agentic systems connecting models to external tools, databases or live data sources multiply that exposure further. The breach Rogoff cites is precisely the kind of incident that emerges from disabling safeguards during testing, even temporarily and even inside a controlled environment.
The Case for Anticipatory Governance
Patching existing law onto AI as incidents accumulate is, in Rogoff’s view, the wrong approach. By the time harm is visible at scale, institutional lag will have compounded the damage. His preferred alternative is governance designed in anticipation of risk: frameworks that create space for development while building in safety, labour and competition considerations from the start. Whether legislatures move at the pace the technology demands remains open. As the EU AI Act’s enforcement timeline shows the political economy of AI regulation rarely favours speed on the governance side.
