Cohere CEO Aidan Gomez argues a few labs shouldn't write the world's AI rules

Cohere co-founder and CEO Aidan Gomez has published an essay on his company's blog arguing that the central question in AI governance is not whether the technology needs rules, but who gets to write them. The piece, which reached the front page of Hacker News, takes aim at a roadmap Gomez says was published this week by Anthropic CEO Dario Amodei, and accuses the largest Silicon Valley labs of seeking antitrust exemptions that would let them set standards for everyone else.

The argument in brief

Gomez's opening concession is that AI needs guardrails, which he calls undisputed, and he says independent review of highly capable systems is a good idea that he supports. His objection is procedural and structural. According to his reading of the proposal, a small group of the most powerful labs based in one country would agree on shared standards and on limits to how fast the technology develops; because it is normally illegal for competitors to agree to limit what they produce, the plan would ask governments for a narrow antitrust waiver, and would then require every other developer to follow whatever the participants settle on.

He notes there is no public comment period, no consultation, and no vote in such a scheme. Adding a few more companies to the table does not fix it either, he argues: the problem is having an exclusive list at all, when the decisions reach every company, government, and citizen who was never asked.

The history Gomez invokes

To make the case, he draws on two regulatory episodes. In 1975 the U.S. Securities and Exchange Commission designated three firms as Nationally Recognized Statistical Rating Organizations for its capital rules, without publishing criteria for how anyone else could qualify. Twenty-five years later there were still only three, Gomez writes, and they went on to rate subprime mortgage securities triple-A ahead of the global financial crisis.

Europe ran a similar experiment in 1985, he says, when car manufacturers lobbied for the Motor Vehicle Block Exemption, an antitrust waiver justified by the claim that complex, safety-critical vehicles required manufacturer control over who could sell and service them. What followed was not safer cars, according to Gomez, and the European Commission needed roughly twenty-five years of reforms to unwind it. In both cases, he writes, the stated goal was safety and nobody set out to build a cartel, yet the result was a market structure that protected incumbents and limited competition.

The technical objections

Beyond process, Gomez argues that a safety regime designed by a few labs will be rigorous only about the risks those labs have already built systems to assess, and quiet about everything else. Risk in existing frameworks gets defined as a function of scale, he writes, which effectively makes the companies with the largest systems the only ones qualified to judge.

He points to open scientific disagreement in the field: how much offensive capability comes from raw model size versus the harness of tools and verification steps wrapped around a model. Smaller models orchestrated well can do things that large models cannot, he argues, and a coordinated swarm of systems presents a different risk surface than a single model. None of that shows up in a regime built exclusively around massive compute thresholds.

He also highlights a promise in the essay that coordination would give developers time for safety work without giving up commercial advantage, and asks whose advantage that preserves. The firms drafting the framework are the firms sitting at the top of the market today, he notes, so a mechanism that slows everyone down while explicitly preserving existing advantage risks turning today's lead into the baseline for competing safely.

What Gomez is not saying

The essay is not an argument against safety rules. Gomez describes Cohere's own systems running inside banks, telecommunications networks, and defense ministries, and says cyber offense is getting cheaper faster than defenses are improving, a gap he treats as genuinely worrying. He also frames his position as pro-science rather than anti-safety: rules should be set based on scientific evidence and open participation, not handed to a small set of commercially aligned executives.

One piece of context readers should weigh: Cohere competes with the labs whose proposal Gomez is attacking, and would be among the developers bound by standards written elsewhere.

Why it matters

The essay reframes a heating-up governance debate from whether to regulate AI to who regulates it, and through what process. If Gomez's historical parallels hold, safety rules written by incumbents can harden into market structure that takes decades to undo, a direct concern for competition authorities already scrutinizing the AI sector. His technical argument, that compute thresholds miss how risk actually emerges from well-orchestrated smaller systems, is a question regulators will have to answer no matter who drafts the standards. And the public split between an enterprise-focused challenger and the frontier labs suggests the industry's sharpest divide right now may not be over capability, but over who controls the rulebook.