In May 1972, Richard Nixon and Leonid Brezhnev signed the agreements known as SALT I after years of negotiations over weapons capable of destroying both countries.
Even with mutual annihilation as an incentive to cooperate, neither side would accept the other’s word alone. Onsite inspections inside the Soviet Union were politically unacceptable, so the agreement relied instead on what it called “national technical means of verification”: Each side watched the other through its own technology, including reconnaissance satellites.
Agreement was desirable. Verification was indispensable. That’s the distinction I keep coming back to when reading Dario Amodei’s “We Must Pace the Frontier.”
Amodei’s essay deserves to be taken seriously. His concern is that artificial intelligence capability is advancing faster than our ability to understand and control it, particularly as AI becomes capable of building the next generation of AI. His response is a three-step plan. First, give independent evaluators ongoing employee-level access to frontier labs to verify safety practices, something Anthropic has said it is committing to unilaterally. Second, get leading labs in democratic countries to align on shared safety standards, with government support where necessary. Third, pursue international agreements, including with China, on the most dangerous capabilities.
I agree with much of the diagnosis. Where I differ is how much weight the second and third steps can bear.
Competition finds a way
Slowing the AI frontier asks companies competing for enormous economic rewards to collectively restrain themselves. Independent verification asks companies to expose more of what they are doing. Pacing asks them to do less of it. Those are very different economic propositions. Independent verification can coexist with a market that rewards speed. Collective restraint asks that market to work against its own incentives.
There is a deeper problem underneath that. Building is what scientists, engineers and entrepreneurs do. We can regulate what gets built, constrain dangerous uses and set consequences for crossing lines. But asking an entire technical field to suppress its own capacity to advance fights economic forces and one of the oldest impulses behind scientific progress at the same time: If something looks possible, someone is going to find out whether it is.
If getting competitors to agree is difficult, getting governments to agree is even harder. Who determines the correct pace of AI development? The United States? China? Europe? Frontier labs? An international organization? And who gets a voice in setting the boundaries when the people affected by these systems are not the people building them? Once there’s an agreement, who determines when someone has crossed the line, and what happens when an entire country decides the cost of slowing down outweighs the risk of breaking the agreement?
This isn’t an argument against international coordination. We should pursue it. We just shouldn’t make human safety dependent on it succeeding, and Amodei’s own essay suggests a genuine global slowdown as the least likely of his proposed outcomes.
The debate broke almost immediately
Within days the essay had split the industry into three camps, and that split is itself useful evidence.
The first camp supported the proposal quickly. Sam Altman committed OpenAI to matching Anthropic’s embedded-evaluator pledge within hours of publication. Elon Musk endorsed the broader call. Google DeepMind’s Demis Hassabis said the essay points toward the right path forward. European Commission President Ursula von der Leyen backed the call to “pace the frontier” and said she would convene the major labs to discuss how Europe could support those efforts.
The second camp accepted the diagnosis and rejected the cure. Meta’s Mark Zuckerberg rejected a coordinated slowdown arguing that competition, legal liability, and independent evaluation already give labs enough reason to build safely, without needing anyone else to set their pace. Investor David Sacks made the same challenge more bluntly: If the companies calling for slower progress believe it is necessary, they can slow themselves down.
The third camp questioned the incentives. Investor Michael Burry called the slowdown talk self-serving, arguing that it could benefit incumbent labs facing growing competition.
We don’t need to decide which of those three camps has it right. That’s the point. Even in the most generous case, where everyone involved is arguing in good faith, the proposal still split the industry inside a couple days. The speed of that disagreement tells us something important: Coordination may be part of a safety strategy, but it cannot be the foundation.
What ‘The process was followed’ actually got us
Aviation is often held up as proof that society can make extraordinarily complex technology safe. It can. It also gives us the Boeing 737 MAX.
Two MAX aircraft crashed in 2018 and 2019, killing 346 people. The U.S. Department of Transportation Inspector General found that Boeing and the Federal Aviation Administration had, in fact, followed the established certification process. The process itself wasn’t enough. FAA guidance contributed to a serious misunderstanding of the flight-control software implicated in both crashes and the agency didn’t fully grasp Boeing’s own safety assessments of that system until after the first crash.
Part of the problem was structural. Through its Organization Designation Authorization program, the FAA had delegated certain certification functions to Boeing. Even a mature, safety-critical regulatory regime can develop a gap between formal compliance and meaningful visibility into the system being regulated.
That’s the lesson for AI. Rules are only as effective as our ability to know whether complex systems are actually following them, and AI makes that problem much harder.
AI creates a verification problem at machine speed
An aircraft is extraordinarily complicated. AI introduces something different: systems whose behavior varies with context, and whose capabilities, tools, permissions and interactions can change rapidly. Agents write code, call APIs, talk to other agents and execute actions faster than a human evaluator could reconstruct what happened after the fact.
METR’s 2026 Frontier Risk Report points at the other side of the problem. Anthropic, Google, Meta and OpenAI gave METR access to their most capable internal models, including raw chains of thought, plus nonpublic information about how those models were being used and monitored internally. That’s a real move away from a world where the company building the technology is also the only source of evidence about whether it’s safe.
But access alone does not establish independence. Who chooses the evaluator? Who sets the standard they’re measured against? What can the evaluator disclose? And what happens when the evaluator and the company disagree? A voluntary assessment is progress, but it is not the same as structurally independent oversight.
And even genuinely independent evaluation runs into a scaling problem: Humans cannot watch this fast enough.
This isn’t only a frontier-lab problem anymore. Deloitte’s 2026 survey of more than 3,200 business and information technology leaders across 24 countries found that only 21% of organizations have mature governance for agentic AI. IBM found that 70% of technology executives say teams across their businesses are already deploying technology faster than IT can track, while only 11% said they were fully prepared for AI-agent deployment at scale.
The available enterprise evidence suggests that deployment is already outpacing many organizations’ ability to govern it.
Governance has to keep up
The scalable answer to increasingly capable AI cannot be an ever-larger number of humans trying to watch it. Likewise, it can’t depend entirely on governments agreeing on the correct speed of progress and successfully enforcing that speed against every company and country whose interests diverge. If AI is acting faster than people can review or intervene, parts of the solution must operate at machine speed too.
That does not mean machines decide what is acceptable. It means humans must be clear about the division of responsibility.
Humans set legitimate boundaries. Governments establish laws and regulations. Organizations determine acceptable risk, define policies and establish consequences for crossing them. The people affected by these systems should not be excluded from those decisions simply because they did not build the technology.
Independent institutions verify. The entities setting policy or selling AI systems should not be the sole judges of whether those systems comply. Evaluators, standards bodies, regulators, auditors and other independent institutions need meaningful access, technical capacity and freedom to report what they find.
Machine-speed systems identify, assess, monitor and enforce. Increasingly, AI systems will need to help discover other AI systems, track what they can access, test their behavior, monitor workflows, identify policy violations, take predefined enforcement actions and preserve evidence of what happened. The objective is not to make another AI system the final authority. It is to automate human-defined rules and make them enforceable in real time in situations where humans cannot supervise continuously.
Humans remain accountable. Accountability does not require a person to perform every act of governance manually. It requires a person, organization or public institution to remain answerable for the rules, the controls, the exceptions and the consequences.
For enterprises, that means governance cannot remain a review process applied to AI after the important technical decisions have already been made. It has to become a first-class part of the technology stack, alongside security, identity, data and observability.
Follow the incentives
There’s already an extraordinary global race to build more capable intelligence. The incentives are obvious: Some of the world’s best engineers and enormous amounts of capital are focused on making models more capable, agents more autonomous and AI more useful.
Governing those systems deserves comparable technical ambition: better evaluation methods, independent verification, continuous monitoring, enforceable policy, stronger standards and infrastructure capable of operating at AI speed.
There should be vigorous competition over how to solve those problems. No single company, lab or technical architecture is likely to have all the answers. More researchers, startups, standards bodies and established technology companies working on governance would be a sign of progress, not fragmentation.
Today, the incentives are lopsided. We reward capability breakthroughs with investment, valuation, customers and geopolitical attention. Governance too often arrives as a compliance cost after deployment.
Closing that gap will take more than policy statements or governance software. Governments should support research into evaluation, monitoring and enforcement alongside capability research. Standards bodies and independent evaluators need access and technical capacity. Enterprises should make governability a requirement of architecture and procurement. And the industry should treat governance as an engineering discipline — not merely a compliance function.
If competition is going to accelerate AI capability, we should create the conditions for competition to improve our ability to govern it too.
Pacing can buy time. It cannot be the end state
Amodei is right that the gap between AI capability and our ability to understand and control it is real.
If frontier labs can coordinate around specific capability thresholds, that may buy us time. Agreements between governments around particularly dangerous applications may reduce some risks. And deeper access for genuinely independent evaluators could give us better evidence about what frontier systems are actually doing.
All are worth pursuing. None should be mistaken for a safety architecture that can stand on its own.
Arms control taught us that agreement is stronger when it can be independently verified. Aviation taught us that even mature regulation fails when the regulator can’t see inside the system it oversees. AI is confronting us with another lesson: Human-speed governance cannot indefinitely keep pace with machine-speed intelligence.
And competition teaches us something too: Incentives work. We have created extraordinary rewards for making AI more capable. We need comparable rewards for making it governable.
A solid plan for the future of AI cannot depend on good intentions. We should not have to decide whether a frontier lab’s motives are altruistic, commercial, or some combination of both. Good governance should work when interests diverge.
That means designing for the incentives that exist today, not the cooperation we hope for.
Pacing may buy time. We should use it to make governance capable of keeping up.
Emre Kazim is co-founder and co-chief executive officer of Holistic AI. He wrote this article for SiliconANGLE.
