Cohere’s Gomez Challenges Lab-Led AI Safety Rules

Cohere co-founder and CEO Aidan Gomez has published a direct challenge to proposals that would let leading AI labs coordinate on safety standards through antitrust exemptions. Gomez argues that AI needs guardrails, but says the rules should be developed openly, internationally and independently—not by the companies being judged under them.

The immediate target is a roadmap Gomez attributes to Anthropic CEO Dario Amodei. Gomez says the proposal would allow a small group of powerful labs to agree on shared standards and limits on the pace of development, then seek a narrow government waiver from competition rules so the coordination is lawful. In Gomez’s account, other AI developers would be required to follow the resulting framework without participating in writing it.

The authority behind the standard

That distinction matters because Gomez is not arguing against external review. He says Cohere supports independent assessment of highly capable AI systems. His objection is to the identity of the people who set review standards, oversee the reviewers and decide who gets a seat in the process.

The dispute is over who writes them, who gets to participate and whose interests the rules are protecting.

Aidan Gomez, Cohere co-founder and CEO

Test what systems can do

Gomez’s alternative begins with a different trigger for regulation. Rules should turn on what a system can do and the conditions in which it is deployed, he argues, rather than on the developer’s size or resources. Under that principle, a dangerous capability would receive the same treatment whether it came from a major lab or a university department.

His case challenges a framework centered chiefly on scale. Gomez says there is real disagreement over whether offensive capability comes from raw model size or from the surrounding system of tools, orchestration and verification. A smaller model used in a coordinated system, he contends, can create a risk profile that compute thresholds alone may miss.

That position is not a rejection of stricter obligations for more capable systems. Gomez proposes a tiered, proportionate approach: systems with more dangerous capabilities, or systems used in particular high-risk settings, would face more demanding tests. The distinction is that the test should follow evidence of harm, not the resources behind the developer.

The framework Gomez wants considered

• An internationally coordinated, public risk framework involving governments, technical experts, critical-sector specialists and researchers who disagree with one another.
• Mandatory disclosures on systems’ construction, purpose, capabilities, risks, mitigations and serious incidents.
• Proportionate independent testing for evidence-based risks, including cyberattacks, synthetic fraud, voice cloning, large-scale manipulation, weapons and critical-infrastructure impacts.
• Open certification and assurance using published criteria, without reviewers being paid by the companies they assess.

A different answer to the compliance problem

The proposal’s practical challenge is institutional. Gomez wants testing capacity funded through public research bodies and existing sectoral risk-management systems, while assurance criteria would be collectively developed and published. Reviewers, he argues, should not be paid by the party under review, and findings should reach the public in some form.

That leaves major design questions open: which institutions would have the expertise and authority to run such a system, and how could an international process reach standards quickly enough for fast-moving technology? Gomez does not claim to settle those questions. His argument is that open certification and limited, evidence-scoped testing are less likely to turn safety compliance into an expense only the largest labs can absorb.

Anthropic has separately said it favors a national AI standard and supported a California measure requiring the largest AI developers to publish frontier-model safety protocols, while exempting companies below a revenue threshold. Gomez’s intervention presses a narrower governance question beneath that wider agreement on oversight: whether large developers should receive a special role in setting the rules that govern the field.

For Gomez, the stakes are concrete. He says Cohere systems operate in banks, telecommunications networks and defense ministries, where failures can carry consequences beyond an individual user. He also warns that AI-enabled cyber offense may be getting cheaper faster than defensive capabilities improve. His conclusion is not that safety can wait, but that strict safety rules need not make incumbent companies the gatekeepers of compliance.