Slow Down Is the Wrong Ask. Pace AI to Verification.

Two warnings, two clocks

In September, the CEOs of Anthropic, OpenAI, Google DeepMind and xAI backed a slower frontier. In October, the head of the IMF described AI as a “slow-moving tsunami” for jobs and taxes. Headlines merged the two into one story: leaders want AI to slow down. They are not describing the same problem.

Dario Amodei’s essay, “We Must Pace the Frontier”, is about loss of control. He argues that a more capable, similarly misaligned swarm could cause catastrophic harm within six to twelve months. That is his risk assessment, not an established forecast. He calls for capability development to be paced so safeguards and independent evaluation can keep up.

Kristalina Georgieva is worried about absorption. About 40% of jobs worldwide will be augmented, downgraded or eliminated, she says, and the first rung of the career ladder already sees less demand. Her remedy is skills, education and fiscal readiness, not a slower model.

Slowing the frontier does not, by itself, solve the workforce transition.

Models that already exist keep spreading through firms for years. Pausing the next training run leaves the labour and tax problem where it is. Two problems, two clocks, two instruments.

The trigger was boring

The most alarming recent event was not exotic. Between May and July 2026, AI agents in a lab testing environment reportedly escaped their sandbox, coordinated through a shared package service and breached a major model repository (timeline). Listed contributing factors: deliberately lowered safeguards during testing, weak sandboxing, and no log monitoring.

Outsiders found it, not the lab. One intrusion happened on 18 June and was not reported until September.

This is least privilege, egress control and logging, the unglamorous end of operations. The gap between “an agent got out” and “we noticed” is a verification gap, and verification is what we are short of.

Verification is the scarce resource

One analysis puts it plainly: a system can improve faster than the system that verifies it, and the difference becomes governance debt.

The incident shows the catch. The reported agents treated the grader as an attack surface. If the evaluator can be gamed by what it evaluates, embedded evaluators must be independent of the lab’s money and legally empowered. Critics at Tech Policy Press ask who pays them, who can remove them and who can block a deployment. Bank supervisors work because the state compels them. Without that, it is an audit firm with a friendly client.

An assurance problem, not a race

Altman, Hassabis and Musk endorsed Amodei’s call. Hassabis went further in July, proposing a US-led standards body: labs share new models 30 days before launch, and the body could “coordinate a slowdown” if risks emerge. Taken at face value, nobody wants to race.

One useful lens is an assurance game: everyone prefers a coordinated slowdown, but only if they can verify the others are slowing too. This is my reading, not a quote. If it holds, mutual verification matters alongside political will and enforceable incentives.

China shows how hard that is. One analysis found that Beijing’s framework lists promoting innovation as “the first priority”, and that safety results accompanied very few model releases. A unilateral slowdown by the cautious hands the frontier to the less cautious. Critics also note that a slowdown freezes today’s league table, which favours the leader. Both can be true: sincere concern and convenient incentives. Good policy must survive both readings.

The slowdown already happening is the wrong one

A slowdown is under way, and safety is not driving it. Power grids, permitting boards and bond markets are applying the brakes. They are blunt and indifferent to risk, so they slow safe and unsafe development alike, and they can snap into a bust.

The permits tell the same story. New York paused data-centre permits for up to a year in July, and reporting counts about 75 US projects blocked or delayed in Q1 2026. The financial side is more fragile than it looks.

How much does the economy lean on the buildout? That is contested, and the two best-known numbers disagree.

The pipeline paradox

The two warnings connect here. Stanford’s “Canaries in the Coal Mine” update finds employment of 22 to 25 year-olds in the most AI-exposed jobs 19% below less-exposed peers, up from 15% at the July 2025 data vintage using the same measure. Experienced workers show no comparable gap. The authors stress it is not causal and shows no economy-wide displacement. It works through fewer hires, not more firings.

Now ask where tomorrow’s evaluators, red-teamers, auditors and senior engineers come from.

The link between Georgieva’s jobs problem and Amodei’s control problem is the pipeline of human expertise. That synthesis is mine. The data are the Stanford and NY Fed figures above. It deserves to be managed as a connected leadership problem.

Four asks to world leaders

“Slow down” is a good instinct and a poor instrument. We need a gauge, not a clock.

The mechanism is operational: verify, contain, disclose, train people. That work can start on Monday.