Cohere’s Gomez: AI needs evidenced standards, not a cartel

Cohere co-founder and CEO Aidan Gomez published an essay arguing that a handful of Silicon Valley labs should not write global AI safety standards behind an antitrust carve-out, and that independent review should rest on open, evidence-based risk rules rather than a cartel.

A lab CEO put Sunday pushback on the record: in Gomez’s reading, Amodei’s pacing plan is an antitrust and cartel risk. The fight is no longer only “slow versus race.” It is who gets to write the rules.

Cohere co-founder and CEO Aidan Gomez published the company-blog essay “Who Gets to Define the Rules for AI?,” subtitled “AI Needs Evidenced Standards, Not A Cartel.” That Cohere post is the filing event. These are Gomez’s Sunday remarks. They are not a Cohere pause, not a granted antitrust waiver, and not a withdrawal of Amodei’s essay.

Gomez’s core question: should a handful of market-dominant Silicon Valley AI companies define safety standards and how fast the technology progresses for the world — while seeking to “bend competition rules”? He writes that those firms are asking to dictate the terms for everyone else. He calls that “a cartel by any other name.”

He agrees independent review of highly capable AI is a good idea. Quoted: “Independent review of highly capable AI systems is a good idea and we support it.” The dispute, he writes, is who writes the standards, who conducts or oversees the review, and whose interests the rules protect. AI needs guardrails, he says. That is not the dispute.

He frames Amodei’s roadmap — published this week — as asking governments for antitrust exemptions so a small set of labs can coordinate shared standards and pace limits, then have governments require other developers to follow, with no public comment, consultation, or vote. File Gomez’s reading of that essay.

Historical analogies Gomez uses as his argument: the 1975 U.S. NRSRO bond-rating designation — nationally recognized statistical rating organization, the official stamp that made a few credit-rating firms the gatekeepers — and the 1985 EU Motor Vehicle Block Exemption. In both cases, he writes, safety-framed structures protected incumbents.

Four pillars Gomez proposes: (1) an evidence-based international risk framework built in the open; (2) mandatory transparency and incident reporting; (3) testing scoped by evidence — cyber, fraud and voice cloning, manipulation, bio/weapons, critical infrastructure — tiered by capability and context, with open certification; (4) real assurance on the model of finance, aviation, and nuclear — criteria published, third parties not paid by the audited party, public findings — and flexibility, against permanently embedded evaluators handpicked by dominant labs.

He rejects extinction “vibes, expressed as decimals” as a basis for binding rules. On the July agent failures (OpenAI–Hugging Face context), he writes the breakdowns were about instructions, the walls around the test, and observability — being able to see what the system actually did — not solved by compute thresholds or more of the same embedded-evaluator setup. This is not a re-file of those July incidents.

On sovereignty, Gomez writes that critical infrastructure needs local, on-prem capable suppliers — systems you can run on your own premises — not rented foreign monopoly interfaces. Competitive diversity, he argues, absorbs failure. A “state-sanctioned cartel” does not.

In closing, he writes that Cohere chooses to “build responsibly… enable sovereignty… work with and listen to lawmakers and safety experts” rather than only talk about slowing.