A Cartel by Any Other Name: Cohere's Aidan Gomez Publishes the Structural Case Against a Lab-Run AI Safety Regime

For a week, the AI safety conversation has been dominated by the builders of the largest models: Dario Amodei’s 3,800-word call to “pace the frontier,” Sam Altman’s openness to slowing down, and reports that OpenAI, Anthropic and Google DeepMind have spent months quietly designing an industry-run standards body. On September 13, the most detailed counter-position arrived — not from a regulator or an academic, but from a rival CEO who builds frontier-adjacent models for banks, telecoms and defense ministries.

Aidan Gomez, co-founder and CEO of Toronto-based Cohere, published “Who Gets to Define the Rules for AI?” — a 15-minute read whose subtitle states its thesis plainly: AI needs evidenced standards, not a cartel. The essay has since ricocheted through the industry, earning coverage from The Times, The Globe and Mail, Bloomberg Television and Barron’s, and giving antitrust skeptics of the slowdown the most articulate voice they have had so far.

The question nobody is asking clearly enough

Gomez opens by conceding the stakes rather than disputing them. He runs a company whose systems are deployed inside banks, telecommunications networks and defense ministries — “among the most high-stakes environments because failures in these sectors can have consequences far beyond an individual user.” He does not downplay risk: “The same capabilities that find vulnerabilities in your code can find them in someone else’s, and cyber offense is getting cheaper faster than defenses are getting better.”

“AI needs guardrails. That is not the dispute and never has been,” he writes. “The dispute is over who writes them, who gets to participate and whose interests the rules are protecting.”

His framing of the core question: “Should a handful of select, market-dominant AI companies from Silicon Valley get to define the rules and safety standards of a generational technology for the entire world? All while simultaneously determining how fast this technology progresses?” His answer is blunt: “A wolf in sheep’s clothing, a cartel by any other name.”

Two history lessons

The essay’s most distinctive move is that it treats “cartel” not as an insult but as a technical term with a documented history — and spends two sections on precedents.

In 1975, the U.S. Securities and Exchange Commission needed reliable bond ratings for its capital rules and designated three firms as Nationally Recognized Statistical Rating Organizations — without ever publishing criteria for how anyone else could earn the same status. These were government-blessed outside evaluators, paid by the very issuers whose securities they graded, sitting behind a barrier the regulator itself had built. Twenty-five years later, there were still only three of them. “Then they rated subprime mortgage securities triple-A and nearly took the global economy down with them.”

Europe ran a parallel experiment in 1985. Car manufacturers lobbied for a sweeping antitrust waiver — the Motor Vehicle Block Exemption — arguing that modern vehicles were complex, safety-critical machines, and that manufacturers therefore needed control over who was qualified to sell and service them. The regulation let manufacturers set standards for premises, equipment and staff training, explicitly in the interest of safe and reliable vehicles. What followed was not safer cars; it took the European Commission roughly twenty-five years of reforms to unwind. Gomez’s lesson: “It is possible you can hold strict safety standards without handing the incumbents a monopoly on meeting them.”

“Nobody set out to build a cartel in either case. In both cases, the stated goal was safety,” he writes. “But the result was a market structure that protected incumbents and limited competition, all under the justification of serving the public interest.”

What he actually objects to in the Amodei roadmap

Gomez is careful to separate what he supports from what he rejects. “Independent review of highly capable AI systems is a good idea and we support it.” The target is the structure around it. Dario Amodei’s September 12 roadmap — which calls for embedded independent evaluators with employee-like access, coordination among frontier labs on safety standards, and government-mediated waivers of antitrust restrictions — is, in Gomez’s reading, a plan in which “a handful of the most powerful labs based in one country would agree on shared standards and the limits to how fast the technology should advance,” then ask governments to “require every other AI developer to blindly follow whatever the participants settle on.”

Three specific criticisms carry the argument:

The participation problem is structural, not arithmetic. “Adding an extra chair fundamentally doesn’t solve the issue. The problem is that there is a list at all, when the decisions being made reach every company, every government, and every citizen who never got asked.” There is no public comment period, no consultation, no vote — “the public will be forced to live with the outcome regardless.”

Scale-based risk definitions entrench scale. Existing frameworks define risk as a function of model size, “which makes the companies with enormous systems the only ones qualified to judge.” But there is genuine scientific disagreement about how much offensive capability comes from raw model size versus the harness wrapped around it. “Smaller models orchestrated well, using tools and verification steps, can do things that large models can’t. A cyber swarm is a completely different risk surface than a single model. None of that shows up in a regime built exclusively around massive compute thresholds.”

The commercial-advantage sentence. Gomez flags one line in the frontier-lab proposals as one “any competition authority would find troubling” — the promise that a coordinated approach would give developers time to do safety work “without sacrificing commercial advantage.” But to whose advantage? “The firms drafting the framework are the firms sitting at the top of the market today. A mechanism that slows everyone down while explicitly preserving existing commercial advantage does not make AI safer. It risks entrenching today’s dominant AI companies by turning their current advantages into baseline for what it takes to compete safely.”

He also reads the entry requirements as a moat by other means: vast computing power, continuous monitoring infrastructure, dedicated security organizations, resident evaluator teams with desks and badges, shutdown architecture, and a pool of “independent” evaluators that is already remarkably small and funded by the same handful of organizations. His summary is caustic: “Convince a government that AI is an existential threat and you can convince it to outlaw your competition.”

The four-pillar alternative

The constructive half of the essay proposes four pillars for a regime Gomez explicitly says should not be written by him either — “nor do I think I should get to make the rules instead”:

1. An evidence-based risk framework. Before anyone mandates testing, publish an agreed account of which harms matter, which capabilities cause them, in which contexts, and when government should step in — built internationally, in the open, “not led by any one nation.” Rules should bind based on what a system can do, “so a dangerous capability is treated the same whether it comes out of a trillion dollar lab or a university department.”

2. Mandatory transparency. Building on today’s model cards: better incident reporting across the development and deployment stack, and “mechanisms to attach real accountability when real harm occurs.”

3. Testing scoped by the evidence. Independent testing of the most advanced systems, but only against capabilities the risk framework has identified as genuinely dangerous — cyberattacks, synthetic fraud and voice cloning, manipulation at scale, weapons, critical infrastructure. Tiered and proportionate, with certification open to every company rather than a designated tier.

4. Real assurance mechanisms. Layered assurance modeled on finance, aviation and nuclear: developers test, customers validate, independent third parties assure, regulators oversee — with third parties “never paid by the party they’re reviewing.”

“Gut feelings, vibes, expressed as decimals”

The essay’s sharpest section takes on the extinction-risk rhetoric that has fueled the slowdown moment — including the researcher who quit a large lab warning of catastrophe, and the senior colleague who put the odds “above ten percent.” Gomez does not question their sincerity. But: “Those numbers didn’t come from any fundamental reality. They are gut feelings, vibes, expressed as decimals, amplified by executives with vested interests and covered by the media for a week as though they were mathematical analyses.”

His technical counter: the real worry is that as systems grow more capable, “the distance between what we asked for and what we actually get becomes harder to notice and more expensive when we miss it.” Whether that adds up to loss of control “is a judgement call, not a finding” — and “an open question of this kind is exactly what a public, contested, evidence-based process exists to work through.”

He points out that the failures reported in July happened inside “the two best-resourced labs in the world, with the largest safety teams” — one with a METR evaluator arrangement being stood up as recently as February. “The proposed remedy is more or less what was in place when it broke.” What failed was instruction quality, test-environment containment, and observation latency — none of which a capability threshold or compute limit fixes. His unglamorous list: mandatory serious-incident reporting, testing against known-exploited gaps, test-time observability standards, and hard isolation for anything wired into hospitals or electrical substations. “A small, poorly specified model sitting inside a hospital is a live risk today, and under a frontier-only regime nobody is even looking at it.”

Why this lands harder than the usual pushback

Two context facts make the essay more than a competitor’s complaint. First, Cohere just closed a reported $3 billion sovereign-AI round and signed a first-of-its-kind transatlantic sovereign AI agreement with Germany’s Aleph Alpha on September 16 — its commercial thesis (governments want capable systems inside their own walls, from replaceable providers) depends on the opposite of a locked-in, Washington-mediated regime. Gomez says so himself: “Critical infrastructure cannot be secured by renting national capability from a foreign monopoly behind a closed interface.” Second, the antitrust concern is no longer hypothetical — the Wall Street Journal’s opinion page argued this week that an AI antitrust exemption “would invite collusion,” and OpenAI itself has privately asked Congress whether a coordinated slowdown is even legal.

The essay closes with the choice as Gomez sees it: rules “written by a group anyone can join and with evidence anyone can check, or by a handful of companies in a room with the door shut.” More voices make the process slower and harder — “some of those voices will say things the rest of us don’t want to hear. That’s the point.”

Whether or not one accepts the cartel framing, the essay has already changed the debate’s shape: the question is no longer only whether to slow down, but who holds the pen when the rules get written — and that question now has a CEO, a history lesson, and a growing chorus behind it.