A dispute over who writes the rules
Cohere co-founder and chief executive Aidan Gomez has entered the debate over artificial-intelligence regulation with a warning that safety standards could entrench the largest developers if those companies are allowed to design the rules themselves. In an essay published on September 14, Gomez said governments should build policy around independent testing, scientific evidence and participation beyond a small circle of US laboratories.
His intervention responds to a roadmap from Anthropic chief executive Dario Amodei. Gomez characterizes that proposal as asking governments to permit competitors to coordinate on shared safety standards and limits on development, then extend the resulting requirements to the wider industry. The Cohere executive supports independent review of highly capable systems but disputes a process in which commercially aligned companies would determine both the standards and the pace of progress.
That distinction matters because safety rules can also shape competition. Tests designed around the technologies, risk models and resources of the biggest laboratories may be difficult for smaller companies or researchers to satisfy, even when their systems pose different risks. Gomez argues that concentrating rulemaking power would narrow the range of scientific views considered and give incumbents influence over the conditions under which rivals operate.
An alternative built around testing
Cohere's proposed direction is not an argument against guardrails. Gomez says AI systems used in banks, telecommunications networks, defence organisations and other critical settings can create consequences well beyond an individual user when they fail. Cybersecurity is one example: the same capabilities that help find weaknesses can lower the cost of offensive activity.
The alternative described in the essay would separate the organisations that develop models from the bodies that evaluate them. Standards would be based on measurable capabilities and evidence, with independent reviewers applying published criteria. A broader group of companies, governments, researchers and affected communities would have a route into the process instead of being asked to accept a framework negotiated by a few developers.
Gomez draws parallels with earlier regulatory systems that unintentionally protected incumbents. He cites the limited group of US credit-rating organisations recognized for regulatory purposes and a European exemption that gave vehicle manufacturers substantial control over sales and servicing standards. His point is that policies introduced in the name of reliability can restrict competition when established firms control access.
The essay is an industry position, not an agreed policy or an independent assessment of Anthropic's proposal. It nevertheless makes the governance question explicit: consensus on the need for AI safety does not settle who should define safety, conduct evaluations or decide which risks count. As governments consider oversight for increasingly capable systems, the design of that process will influence both public protection and the structure of the AI market.
