Cohere CEO Criticizes AI Companies' Antitrust Exemption Demands: "A Few Companies Should Not Monopolize the Rules"

Cohere CEO Aidan Gomez criticized attempts by a small number of Silicon Valley AI companies to define global AI rules and safety standards. While acknowledging the need for AI guardrails, he pointed out that it is unjust for rule-making to be monopolized by a few companies and for attempts to limit competition through antitrust exemptions.

Monopoly Structures Similar to Past Cases

Gomez cited the 1985 lobbying case by European automobile manufacturers, warning that structures have been formed to protect the market dominance of incumbent players under the guise of safety. He criticized the roadmap recently proposed by the Anthropic CEO as a request to the government for antitrust exemptions under the pretext of safety, which he argued is equivalent to imposing rules without input from other developers and society.

Limitations of Scale-Centric Regulatory Frameworks

Existing safety frameworks define model scale as a function of risk, assuming that only companies with massive systems are qualified to make judgments. Gomez pointed out that large-scale compute threshold-based systems fail to capture new risk surfaces such as the tool-use capabilities of well-designed small models or cyber swarms. He also analyzed that while proponents claim a coordinated approach does not sacrifice commercial advantage, it effectively locks in the advantages of current market leaders and limits competition.

Proposals for Better Rules

Gomez proposed the following practical ideas for a democratic legislative process:

• Evidence-based risk framework: Publicly agree on harms, causes, and points of government intervention, emphasizing international cooperation involving engineers and experts from diverse fields. Regulatory application criteria should be based on the system's 'capabilities' and 'deployment location,' not the developer.
• Mandatory transparency: Serious incidents must be reported to serve as lessons for the entire field, and standards for test-time observability or logging should be established to detect bad behavior early.
• Core infrastructure isolation: Systems connected to critical infrastructure, such as hospitals or substations, must be isolated, for example, on-premises. Even a small model deployed in a hospital poses an immediate risk, so regulation should be based on the reach of the system, not the size of the company.

The Gap Between Fear Marketing and Actual Risks

Gomez argues that the intention to ban competitors by defining AI as an existential threat does not create a safer world. He criticized figures such as the probability of human extinction from superintelligent systems as being based on intuition rather than fundamental reality. Instead, he offered practical alternatives to address ongoing harms such as voice cloning and automated decision systems.

Ultimately, Gomez emphasized that AI rules should be established through an open, verifiable, multilateral process involving academia, civil society, small labs, and open-source developers, rather than by a closed group of a few companies.