On September 12, 2026, Anthropic CEO Dario Amodei published "We Must Pace the Frontier." He argues that the AI industry should deliberately slow frontier capability gains so that safety work can catch up. In his words, "if slowing down bought us even an extra year or two before models reach critical levels of capability, and we used that time to advance alignment, we could greatly reduce the risk that something goes seriously wrong." Within a week, governments had split along predictable lines. That split matters more than the essay itself.
What the plan actually proposes
The essay sets out three steps. First, each frontier company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators. Second, frontier companies in democracies agree common safety standards, which Amodei says will need government support because of the legal risks of industry coordination. Third, democratic governments try to coordinate with authoritarian states, while conceding that verification will be hard.
Anthropic committed unilaterally to step one. Per the essay, evaluators get office access, badges and company laptops. Their permissions are comparable to those of internal risk teams. They can publish findings without Anthropic's editorial control, apart from security and legal redactions. OpenAI's Sam Altman said his company would adopt the same evaluator-access proposal, according to MediaNama's roundup.
The strongest case for pacing
The case for pacing deserves a fair hearing. Capability gains are arriving faster than the tools for measuring and controlling them. Amodei warns of loss of control, cyber and bio misuse, and economic disruption, and notes that commercial competition can produce a race to the bottom. In her September 16 State of the Union speech, Ursula von der Leyen said that models now in development could enable hacking at a level Europe had not anticipated, Euronews reported. If a race dynamic is pushing labs to ship before they understand their systems, a coordinated pause could be cheaper than an accident.
A fractured government response
The reaction shows how hard step two and step three will be.
- European Union: Von der Leyen said "CEOs of the most advanced companies tell us that it is time to slow down on the self-recursive models." She committed to work with Canada, the UK and others on "model evaluation, verification, early-warning, AI security." She also said she would convene the leading labs.
- United States: President Trump rejected a slowdown and framed AI development as essential to competition with China.
- Germany: The digital ministry said it does "not consider halting development to be a viable option for Europe," citing digital sovereignty.
- China: The foreign ministry called for an "open, inclusive and benevolent approach to AI" and warned that fomenting threats and confrontation would disrupt global AI governance, while the state-run Global Times called the essay a "Cold War playbook" for the AI sector.
- Spain: Digital minister Oscar Lopez compared the debate to nuclear non-proliferation agreements.
The positions are not random. Each government's stance tracks its exposure to the race. Countries that lead in frontier development see a slowdown as a handicap. Countries that depend on imported models see it as a chance to catch up, or as a threat to their sovereignty. A one-to-two-year pause that binds only some players is not a pause, because the others simply pass them.
Where the proportionate approach lies
This is the core problem with a blanket slowdown. Capability is hard to define and hard to verify, and compute alone is a crude proxy. A broad, politically negotiated pacing regime would likely chill open research, university labs and smaller developers long before it touched the best-funded frontier programs. It would also raise the barrier to entry, which favours incumbents.
The better instrument is the one Amodei starts with: independent, evidence-generating evaluation. Institutions for it already exist. The UK's AI Security Institute says it tests leading systems before and after launch and has pre-deployment access to leading AI models. Its focus areas include cyber misuse, safeguard effectiveness, alignment reliability and autonomous capabilities. In the EU, providers of the most advanced general-purpose models are already legally obliged to notify the AI Office, and the Commission's enforcement powers, including fines, took effect on August 2, 2026. Building on these institutions is more credible than inventing a new pacing authority. It also lets regulators tie any restrictions to observed risks, not to a calendar.
The self-regulation objection
Dave Karpf, writing in Tech Policy Press, raises the most serious objection. He argues that companies cannot set the terms of their own oversight, and he points out that Anthropic is itself preparing an IPO. He writes: "The cigarette industry also favored 'sensible and targeted' regulations, so long as it got to set the rules for itself." He compares embedded evaluators to the Arthur Andersen–Enron relationship, in which the auditor's business depended on the audited firm.
The critique has force, but it argues for design, not for dismissal. Evaluator access is worth having only if evaluators are independent of the firm that pays for them and are free to publish. Amodei's publication-rights commitment is a start, but a voluntary promise can be withdrawn. That is the case for putting these terms into law or regulator-run programmes, with common standards, statutory access rights and penalties for obstruction. It is not a case against the idea.
What to watch
The practical test comes in the next few months. Watch whether other labs match Anthropic's access terms and whether evaluators' first reports are actually published. Watch also whether von der Leyen's proposed talks with labs produce binding evaluation protocols, or a voluntary code that lapses under competitive pressure. Pacing that depends on unanimity among rival states will not hold. Verified evaluation, backed by law and open to any credible evaluator, can be adopted by one jurisdiction at a time. Its success in doing so would show whether a slowdown is needed at all, or whether better measurement is enough.