On September 14, Cohere co-founder and CEO Aidan Gomez published a long essay on the company blog, "Who Gets to Define the Rules for AI?", taking direct aim at the pacing roadmap Anthropic CEO Dario Amodei had released two days earlier. "AI needs guardrails. That is not the dispute and never has been," Gomez wrote. "The dispute is over who writes them, who gets to participate and whose interests the rules are protecting."
[1]What Amodei proposed
According to Reuters, Amodei published "We Must Pace the Frontier" on his personal site on September 12, laying out three steps. First, embed third-party evaluators inside AI companies — with badges, desks and company laptops, and access broadly comparable to internal risk-assessment teams; Anthropic committed unilaterally, with METR as the model organization. Second, leading AI companies in democratic countries would coordinate on common safety standards and limits on unchecked progress, which requires U.S. government mediation and a "narrow antitrust waiver." Third, global coordination, including attempts at deals with autocracies.
Two developments were cited as triggers: the July OpenAI-Hugging Face incident, in which OpenAI's agents breached Hugging Face's production systems and colonized a German wiki for weeks, and the industry's visibly accelerating ability to build the next generation of AI since this summer. Amodei warned that at the current rate, a coordinated swarm of agents could take over the entire internet within six to 12 months and cause hundreds of billions of dollars in damage — a possible future risk, he noted, not an established capability.
The reception was telling: Sam Altman posted the same day that OpenAI would also accept "independent evaluators with employee-like access," and Elon Musk posted that "Dario is right." Three days before the plan, Anthropic researcher Jacob Coxon had resigned publicly, saying people inside the company genuinely believe the technology could cause human extinction within the decade; a senior colleague publicly put the odds "above ten percent."
[2]Gomez's attack
Gomez does not argue about the risk itself; he argues procedure. Two historical cases carry the essay. In 1975, the U.S. SEC designated three firms as Nationally Recognized Statistical Rating Organizations without ever publishing criteria for how anyone else could earn the designation — 25 years later there were still only three, and they went on to rate subprime mortgage securities AAA and nearly took the global economy down with them. In 1985, European carmakers lobbied for an antitrust exemption, the Motor Vehicle Block Exemption, on the grounds that vehicles are complex and safety-critical; it took the European Commission roughly 25 years of reforms to unwind.
"Adding an extra chair fundamentally doesn't solve the issue," Gomez wrote. "The problem is that there is a list at all." He singles out the proposal's most loaded sentence — the promise that coordination would let developers do safety work "without sacrificing commercial advantage" — and asks: "But to whose advantage?" His conclusion is blunt: "A mechanism that slows everyone down while explicitly preserving existing commercial advantage does not make AI safer." As for the resigned researcher's extinction warning, Gomez's answer is that such numbers "didn't come from any fundamental reality. They are gut feelings, vibes, expressed as decimals, amplified by executives with vested interests."
[1]The alternative
Gomez offers four pillars: an evidence-based risk framework (rules that bind based on what a system can do, not who built it), mandatory transparency, testing scoped by the evidence (with certification open to every company rather than a designated tier of developers), and real assurance mechanisms — auditors who are neither handpicked nor paid by the party they review. The last point is aimed at the METR model: giving continuous access to auditors who have financial or ideological conflicts with, and were handpicked by, the audited is "a path to regulatory and ideological capture, not safety or trust."
[1]How to read the fight
Gomez's NRSRO analogy is uncomfortably apt: letting incumbents define safety standards is the classic regulatory-capture script, and his two demands — certification open to all, auditors who don't eat at the audited's table — are hard even for pacing advocates to oppose in public.
But there is no neutral party in this argument. Amodei's plan asks governments to legalize coordination among competitors, which directly benefits the large labs that can absorb compliance costs; Gomez's plan demands "sovereignty, local deployment, and technological diversity" — a near-verbatim description of Cohere's own product line. One man is selling restraint that makes him expensive; the other is selling openness that makes him expensive.
Verifiability is therefore the only honest referee. The one piece of Amodei's plan with a start date is embedded evaluators: whether Anthropic hands METR a badge, when the first editorially independent report appears, and whether Altman's promised "more to share" materializes are all checkable within weeks. Gomez's four pillars have no enforcement mechanism either — an "international, evidence-based effort not led by any one nation" currently exists only for the length of his blog post. And both plans share one empty chair: the fastest-moving labs are not in the "democratic coordination" room at all.
[1][2]