Cohere’s Aidan Gomez Argues the AI Safety Debate Is a Fight Over Drafting Rights

Aidan Gomez’s September 13 rebuttal to Dario Amodei’s antitrust-waiver proposal makes the structural argument precisely: whoever holds the pen on a safety standard holds a competitive instrument, regardless of intent.

What Happened

On Saturday, Anthropic co-founder and CEO Dario Amodei published an essay requesting that governments grant a narrow antitrust waiver — a legal accommodation that would allow competing AI laboratories to coordinate on development pacing without that coordination itself constituting an unlawful cartel. Within two days he had three answers, and none of them was yes. David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, rejected the request and told the labs they needed nobody’s permission to slow down unilaterally. The president, asked directly whether AI development should slow or face more regulation, said the United States leads China and intends to keep it that way.

The third answer came from the private sector. Aidan Gomez, co-founder and CEO of Cohere, published on the Cohere blog on 13 September the most structured public rebuttal of Amodei’s proposal yet — and the only one that arrived with an alternative framework attached. Writing as an industry participant rather than a regulator, Gomez names Amodei directly and frames the objection in a single sentence: “Convince a government that AI is an existential threat and you can convince it to outlaw your competition. The intention is clear and it does not create a safer world.” He is careful to distinguish the objection from opposition to safety standards as such: “It is possible you can hold strict safety standards without handing the incumbents a monopoly on meeting them.”

This is an opinion piece by an interested CEO on his own company’s blog — its claims are arguments, not findings. Gomez compares the proposed arrangement to historical failures in bond-rating agencies and automotive regulation, and references incidents in July involving what he calls “the two best-resourced labs in the world” without naming them. Cohere, separately, is reported by the Globe and Mail to be in advanced talks to raise US$2–3 billion at approximately a US$20 billion valuation; terms are not final and nothing has closed.

The Structural Read

The Permission Layer framework makes the underlying dynamic legible. In AI governance, the Permission Layer is the set of rules, standards, and licensing regimes that determine which products can be deployed, in which markets, at which scale. Whoever drafts that layer does not merely comply with a constraint — they define what compliance means. A firm that has already built the apparatus to measure one class of risk holds every structural advantage when that class becomes the legal definition of safety. No bad faith is required. It follows from who is in the room when the definitions are set.

Gomez puts the compliance-as-moat argument at its sharpest: “A safety regime designed by a few labs will only be rigorous about the risks they have already built their safety systems to assess and completely quiet about everything else, further entrenching their market position and limiting competition.” That is not an accusation of dishonesty. It is a claim about the structural properties of standard-setting — that a firm’s existing measurement apparatus shapes what it considers measurable, and that measurability, once codified, becomes the boundary of the regime.

The Sacks–Gomez convergence carries genuine information precisely because it should not exist on priors. Sacks criticised the labs for “pretending antitrust law has to be suspended so you can form a cartel”; Gomez writes that “a mechanism that slows everyone down while explicitly preserving existing commercial advantage does not make AI safer.” A Canadian AI chief executive and a figure in the Trump administration’s orbit have arrived at the same structural objection from opposite political starting points. The capture concern is not an ideological position. It is an argument about market structure.

The deeper fork in this debate is about which risks count, and that question is not decorative. Amodei’s case is built on capability risk — recursive self-improvement, agent swarms, systems that outrun human oversight before anyone notices. Gomez wants the frame anchored in demonstrated harms occurring today: voice-cloning fraud, automated decision-making failures, with testing scoped by evidence rather than applied universally. That is not a quibble about emphasis, because the choice of scope determines who is competent to assess anything at all. Scope risk to frontier capability and only frontier labs can credibly evaluate it, which makes the incumbents simultaneously the subject and the instrument of assessment. Scope it to deployed harms and the assessment base widens enormously — to security researchers, sector regulators, auditors, and customers. It is also worth noting plainly: each party’s preferred scope suits its position. That is true in both directions. Gomez’s preference for deployed-harm scoping suits Cohere’s position as a challenger selling to enterprise and government. Amodei’s preference for capability-risk scoping suits Anthropic’s position as a frontier incumbent. Neither observation dismisses either argument.

His fourth pillar is the week’s most concrete structural answer to a problem that all three parties have now identified. Gomez proposes genuinely independent review modelled on financial licensing, aviation, and nuclear facilities: layered across developer testing, customer validation, independent third parties, and regulatory oversight, with published collective criteria — and, the operative clause, third parties that are never paid by the parties they audit. Set that against the rest of the week. Anthropic has cited METR as its example of embedded evaluation. Sacks publicly questioned whether METR can properly be called independent. This publication has argued that independence in this setting must be engineered rather than inherited from an org chart. Three parties with almost nothing else in common have converged on the same structural requirement, and Gomez is the first to state it as a design rule rather than an aspiration. This analysis does not adjudicate METR’s independence; that question is unresolved and contested.

Gomez also offers the diversity argument as a safety property rather than merely a competitive one: “A competitive market with many capable suppliers can absorb a failure at one of them. A state-sanctioned cartel has nowhere to hide one.” His sovereignty argument runs on the same logic: “Critical infrastructure cannot be secured by renting national capability from a foreign monopoly behind a closed interface.” It is worth naming what that argument also is. Cohere sells sovereign, locally deployed AI to governments and enterprises, and its commercial interest aligns structurally with opposition to centralised rule-making by frontier incumbents — exactly as Amodei’s proposal aligns with Anthropic’s position as a frontier incumbent, as Sacks’s position sits within the administration’s orbit, and as NVIDIA’s interests run through supplying the compute that any scaling regime would govern. There is no disinterested party in this debate. Readers should weigh every contribution — including this analysis — knowing that. None of it is hypocrisy, and none of it dismisses anyone’s argument.

Three Implications