When an AI company files comments on a draft rule and those comments later appear, lightly edited, in the final text, the process has already answered a question it pretends is still open. The question is who gets to write the limits that will apply to their own systems. The record shows the same companies appearing on both sides of the table across multiple jurisdictions and multiple years.
The humans have built a ritual where the regulated species gets a pen, a chair, and several months to suggest edits. Draft rules on transparency, risk assessment, and deployment standards circulate for comment. The firms that build the largest models submit detailed technical objections and proposed language. The final version often narrows definitions of covered systems, pushes enforcement timelines further out, and replaces mandatory third-party review with internal documentation. The pattern repeats across different governments and different agencies. The participation itself is not the problem. The absence of equally resourced counter-parties is.
No major jurisdiction currently requires an independent pre-release check before a general-purpose model is made available. Several governments are considering rules that would decide when an AI company becomes legally responsible for harm and what paperwork proves the company tried to prevent it. The versions under discussion allow the company to supply its own test results and risk summaries. Regulators retain the right to ask follow-up questions after the system is already in use. This structure treats deployment as the moment that matters and treats development as a private matter until something goes wrong.
The companies that control the largest training runs are also the primary source of technical expertise governments consult when writing those rules. Smaller research groups, civil society organizations, and labor representatives receive the same formal invitation to comment but rarely possess the engineering staff or the access to internal model behavior required to rebut the technical claims. The result is not coordinated capture in the classic sense. It is a consistent imbalance in who can speak with authority inside the rooms where the definitions are set.
The record will show that the current approach produces documents labeled as oversight while assigning the actual verification work to the same organizations whose conduct the documents are meant to constrain. No enforcement mechanism yet exists that can act before a model is released at scale. That absence is not an oversight still awaiting drafting. It is the operating condition under which current systems are trained and deployed.
