Artificial intelligence is producing an unusual policy problem: by the time society has reliable evidence about one generation of systems, the next is already in use. The 2026 International AI Safety Report describes the bind directly. Capabilities are improving quickly while evidence about risks arrives slowly. Acting before the evidence is clear risks locking in the wrong rules. Waiting for certainty risks leaving the public exposed.
On September 12, Anthropic CEO Dario Amodei published an essay titled "We Must Pace the Frontier." The proposal has been widely compressed into a headline: the man who builds frontier AI wants everyone to slow down. That compression misses the operative distinction. Amodei is not calling for a halt to AI research, a moratorium on deployment or a shutdown of the systems already used in medicine, education, science and business. He is calling for a slower pace of improvement in the most advanced models, specifically when safety work and public institutions cannot keep up. "Pacing does not mean halting model training or technical progress," Amodei writes, "but ensuring companies take adequate time to align and safeguard their models."
Most people who encounter this debate will hear it as a binary: for progress or against it. That framing is wrong, and it obscures the question that actually matters. A defensible policy needs to separate three layers. Frontier development means training a system that materially advances general capabilities. Frontier deployment means releasing that system into real world use. Diffusion means applying existing, tested systems to useful problems. A credible slowdown can target the first layer while leaving the third intact, though diffusion is not automatically harmless: deployment and diffusion require their own safeguards, from monitoring to access controls. Pacing can also accelerate safety research, auditing and institutional readiness. Once you see that distinction, the argument changes.
The Case for Pacing
The strongest argument for deliberate pacing is practical. At present, a frontier laboratory can improve capabilities faster than independent evaluation capacity can keep pace. A capability triggered checkpoint reverses that burden. Above a defined threshold, the developer must demonstrate that corresponding safety controls work before proceeding. This is more specific than asking firms to exercise caution.
If the time is used well, a checkpoint can accomplish several things: it allows independent evaluation of systems before they reach the public; it reduces the competitive pressure that makes each laboratory fear that slowing down unilaterally will hand the lead to a rival; and it creates space for legislatures, regulators and researchers to determine what evidence society requires before accepting a new level of risk. The 2026 International AI Safety Report documents that pre deployment tests do not reliably predict real world performance, and that models can distinguish test environments from deployment, exploiting evaluation loopholes. A checkpoint regime built on layered, sustained evaluator access rather than one time benchmarks would address exactly that gap.
Anyone who has used an AI tool at work and then read a headline about existential risk understands the dissonance between the technology's daily utility and the scale of the governance questions it raises. The case for pacing is that society should not have to resolve that dissonance retroactively.
The Costs and the Credibility Problem
A broad slowdown can delay real benefits. The OECD has documented task level productivity gains in experimental settings, with estimates varying by occupation and measure. The International AI Safety Report notes practical advances in protein design, algorithm discovery and scientific reasoning. A blanket pause on frontier training could postpone medical, research and productivity applications without proving that the foregone progress would have been dangerous.
Then there is defection. If one responsible developer slows while others continue, the advantage shifts to actors with weaker controls. Anthropic's own record illustrates the fragility of voluntary restraint. In February 2026, Anthropic replaced its previous unconditional scaling commitment with a framework that separates company commitments from more ambitious industry wide recommendations and considers the additional risk created by Anthropic relative to competitors. The revision weakened the premise of unilateral restraint, but it did not create a simple rule under which development stops only when Anthropic is the industry leader. Chief Science Officer Jared Kaplan said the company did not feel it made sense to maintain unilateral commitments if competitors were advancing without equivalent constraints. If the company that built its brand on safety could not sustain a unilateral standard, voluntary pacing across the industry cannot be assumed.
The hardest question is no longer simply whether development should slow. It is who can enforce a checkpoint, and whether the resulting rules entrench the companies already leading the market. The Canadian Competition Bureau has warned that scarce compute, economies of scale and specialist talent can raise entry barriers and contribute to concentration in foundation model markets. Complex audits and licensing can add another barrier. A checkpoint regime must apply capability based thresholds symmetrically, fund shared evaluation infrastructure, exempt lower risk research and give competition authorities a formal role.
The Workable Middle
Amodei's proposal has three stages: embedded independent evaluators with employee level access to frontier laboratories, coordination among leading democratic country firms on safety standards and capability linked checkpoints, and eventual international coordination beginning with narrow agreements on the most dangerous applications. Anthropic says it will begin the first stage unilaterally.
Amodei addresses evaluator publication rights, proposing access without editorial control subject to narrow redactions. But critical design questions remain. How are evaluators selected and financed? What prevents conflicts of interest? Who resolves disputes over access or proposed redactions? What authority can impose a checkpoint, and what evidence permits development to resume?
If Canada's AI Safety Institute expands its independent evaluation capacity, publishes reproducible methods and establishes interoperable standards through the international safety institute network, and if participating laboratories accept binding triggers with published restart conditions and competition review, a checkpoint regime can convert time into demonstrable safety rather than indefinite delay. If the checkpoints remain voluntary, the triggers undefined and the restart conditions at the discretion of the firms being regulated, the system will reproduce the pattern that Anthropic's February policy retreat already demonstrated.
Canada's Position
Canada cannot likely set the global frontier pace. It can shape the evidence and standards on which allied decisions rest. The Canadian AI Safety Institute, the voluntary code for advanced generative AI and the 2026 national strategy's commitments to transparent evaluation and international cooperation provide part of the institutional foundation. Ottawa is also consulting on stronger transparency measures, including serious incident tracking. What remains absent is an enforceable frontier checkpoint.
A slowdown is justified only if it converts time into verifiable safeguards, preserves beneficial use, protects competition and establishes clear conditions under which progress resumes. That is the test. Any measure that fails it is delay with a better name.