Good intentions are not a verification system.
A company can say it is acting responsibly. Famous founders can stand behind the same promise. But the reputation of the people signing a pledge is not the same as the trustworthiness of the pledge.
If a threshold for slowing AI development exists for public safety, competitors and newcomers should be able to understand it. When should work stop? What evidence permits it to resume? Who approves an exception? Without visible answers, safety becomes discretion rather than a rule.
A responsible brand does not merely ask to be trusted. It makes its conduct verifiable even by people who do not trust it.
Who does an independent evaluator answer to?
Anthropic CEO Dario Amodei has proposed third-party evaluators with continuing access inside frontier labs: access to verify safety practices and incidents, and the ability to publish findings without company editorial control, subject to limited security redactions.
That is a useful beginning. But independence does not come from the label. Who selects and pays the evaluator? Can it publish an unfavorable finding? Is the same test available across labs? Can evaluators explain their judgment after leaving a company? Those questions determine whether evaluation is oversight or theater.
Safety rules must not become an entry barrier.
The OECD argues that open and contestable AI markets support innovation. It also warns about concentration in compute, data and skills, vertical integration and first-mover advantages. The need for safety standards and the risk that those standards protect incumbents can both be true.
Rules should therefore attach to capabilities and risks, not company names. Comparable risks need comparable thresholds. Evaluation methods and restart conditions should be visible. Smaller entrants need workable access to evaluation. If safety becomes a license to eliminate competition, society may trade risk reduction for the loss of meaningful choice.
Four things the public should be able to verify.
- Does the pause threshold follow model capability and risk rather than the identity of the company?
- Can evaluators access the needed evidence and publish unfavorable findings independently?
- Are incidents, corrective actions and the basis for restarting recorded?
- Are the rule-maker and the body hearing challenges meaningfully separate?
The OECD due-diligence guidance for responsible AI does not stop at a policy statement. It calls for identifying impacts, preventing and mitigating harm, tracking results, communicating and providing remediation when appropriate. Trust looks less like a pledge and more like a record that can be inspected over time.
Scope and limitations.
This essay does not conclude that any company violated the law. The current lawsuit has produced no judicial findings or judgment, and the governance tests proposed here are the author’s analysis. Independent evaluation and public rules cannot remove every AI risk. As technical capabilities and market structures change, the standards themselves need public review.