Cohere co-founder and CEO Aidan Gomez has challenged proposals that would let a small group of leading AI laboratories coordinate safety standards with an antitrust exemption. His central argument is not that advanced AI needs fewer guardrails, but that the companies being regulated should not also get to write the rulebook.
Cohere Watch analysis
What happened
In a policy essay published by Cohere, Gomez responds to a roadmap from Anthropic CEO Dario Amodei. He says he supports independent review of highly capable systems, but objects to dominant laboratories agreeing shared standards and limits that other developers would then have to follow.
Gomez argues that such an arrangement could preserve today’s commercial hierarchy while narrowing the scientific debate about risk. That is his assessment, not an established finding about the proposal’s motives or eventual effects.
What changes
- Define harms in public
Gomez wants published evidence linking specific AI capabilities to specific harms, contexts and thresholds for government action. - Make the process international
He proposes a multi-country effort involving technologists, policymakers, critical-sector experts and researchers who disagree with one another. - Regulate capabilities, not corporate identity
The same dangerous capability should face the same rules whether it comes from a giant laboratory or a university team. - Require meaningful transparency
Developers should disclose intended uses, capabilities, tests, mitigations and serious incidents across the parts of the deployment stack they control. - Test where evidence points
Independent assessments should target demonstrated risks such as cyberattacks, fraud, voice cloning, manipulation, weapons and critical-infrastructure harm.
Why it matters
The argument exposes a genuine tension in AI governance. Frontier laboratories have the expertise, infrastructure and alarming front-row view needed to identify emerging risks, but rules built around their own systems may also become entry requirements that only similarly enormous companies can satisfy.
Gomez also makes a technically useful point: compute is an imperfect stand-in for capability. A smaller model equipped with tools, orchestration and verification can present a different risk surface from one large model sitting politely in a benchmark harness.
Our read
Gomez lands the important distinction: opposing industry-written rules is not the same as opposing safety regulation. Capability-based standards, transparent evidence and independent public testing offer a more defensible starting point than asking the current leaders to form a safety committee and trusting competition to survive the seating plan.
His critique still needs to meet the same standard it demands. Policymakers should require evidence showing which governance designs reduce real harms without creating a compliance moat, then publish the disagreements rather than polishing them away.
What to watch
- Whether Anthropic answers Gomez’s competition and governance objections directly.
- How any proposed antitrust exemption defines membership, oversight and public accountability.
- Whether regulators adopt capability tests that cover tool-using and orchestrated smaller models.
- Who funds independent testing and whether its methods and results are published.
Discussion spark: Who should set advanced-AI safety standards, and what governance structure could keep both laboratory influence and regulatory capture in check?
Sources and evidence
- Source update (13 September 2026, 21:46 UTC)
not affiliated with or endorsed by Cohere