Aidan Gomez, co-founder and chief executive of Cohere, published an essay on September 13 arguing that the safety framework Anthropic CEO Dario Amodei proposed would hand a small group of Silicon Valley labs the power to set global AI rules while shielding themselves from competition. Gomez does not dispute that frontier systems need oversight. His argument is about who gets to write the rules and who benefits when they do.
Cohere sells enterprise AI systems to banks, telecoms, and defense ministries, which makes Gomez a direct competitor to Anthropic and OpenAI, not a neutral observer. His essay reads as opinion and industry self-interest as much as critique, and readers should weigh it that way: a smaller lab arguing against rules that would favor the largest ones is also arguing for its own market position.
Gomez’s central claim is structural. Amodei’s roadmap, in Gomez’s reading, wants antitrust exemptions from governments so a handful of dominant labs could coordinate on safety standards and pacing, something normally illegal among competitors. Gomez frames that request as a bid to convert safety language into a barrier that keeps smaller developers out of the conversation entirely.
To make his case, Gomez reaches for two historical parallels rather than AI-specific evidence. He points to the Securities and Exchange Commission’s 1975 decision to designate three credit-rating agencies as the only recognized evaluators, a structure that stayed effectively closed for a quarter century before those same agencies handed triple-A grades to subprime mortgage bonds in the run-up to the 2008 financial crisis. He also cites the European Union’s 1985 Motor Vehicle Block Exemption, which let carmakers set their own qualification standards for dealers and mechanics and took roughly 25 years of regulatory reform to unwind. Both cases, in his telling, started as safety measures and hardened into protections for incumbents.
Gomez proposes an alternative built on four elements: an internationally developed risk framework that classifies harms by capability rather than by company size, mandatory transparency requirements beyond today’s model cards, testing scoped to specific dangerous capabilities rather than blanket compliance exercises, and assurance mechanisms with evaluators who are not selected or funded by the companies they review. He argues rules should bind based on what a system can do, so a dangerous capability gets identical treatment in a university project and in a large lab.
He also pushes back on recent public warnings from researchers who left major labs predicting a meaningful probability that superintelligent systems could cause catastrophic harm within a decade. Gomez calls those probability estimates informed judgment rather than measured findings, and argues that framing the debate around existential risk pulls attention away from present-day harms he says are easier to document, including AI-enabled voice cloning used in financial fraud.
The essay lands as regulators in multiple jurisdictions are actively drafting AI governance frameworks, which gives Gomez’s intervention real timing even setting aside its self-interest. Gomez’s essay and Amodei’s roadmap are both best read as competing position papers from companies with a stake in how the rules land.
For operators building on frontier models, the practical signal is that the shape of AI regulation is still contested among the vendors themselves, not settled. Teams making multi-year infrastructure bets should watch which framework, capability-based testing or lab-coordinated self-regulation, gains traction with actual legislators over the next two quarters, since that outcome will determine whether smaller model providers remain viable alternatives to the largest labs.
Based on an essay by Aidan Gomez, co-founder and CEO of Cohere, published on the Cohere company blog on September 13, 2026.
