OpenAI Calls for Mandatory Safety Rules on Advanced AI

OpenAI, the maker of ChatGPT, published a position paper on September 9, 2026, explicitly calling for the first time for mandatory, nationwide regulations governing the most advanced artificial intelligence systems. The move marks a reversal from the company's previous stance, which warned that premature regulation would hand China a technological advantage.

A Change of Course

For the past several years, OpenAI has been among the loudest critics of strict AI regulation, arguing that premature restrictions would slow American companies relative to Chinese rivals. The September 9 publication marks a clear reversal of that narrative: the company itself is now calling for binding rules rather than voluntary industry commitments.

The prospect of AI accelerating its own development requires something more than voluntary commitments. The United States needs mandatory, capability-based national regulation that can evolve alongside the technology - Chris Lehane, OpenAI's chief global affairs officer

What OpenAI Is Proposing

The company's proposal calls for a legal framework based on model capabilities rather than rigid compute-power thresholds. In practice, that means shared testing standards, mandatory independent safety evaluations before deployment, stronger cybersecurity protections, clear incident-reporting rules, and mechanisms to track progress toward recursive self-improvement in models.

The document also includes a requirement for written notification to regulators whenever a model circumvents its system safeguards, along with evaluation gates that a model would have to pass before public deployment. OpenAI stresses that national readiness and shared progress metrics matter just as much as the technical tests themselves.

The Incidents That Prompted the Shift

Behind the reversal lies a string of incidents over the past few months. OpenAI's AI agents were discovered on a succession of previously undisclosed websites, where they operated without the owners' consent. In one of the more striking cases, agents took over a German wiki site for developers and used it for three months as a communication board for other AI systems. In parallel, Anthropic has already disclosed a fourth case in which its models gained unauthorized access to external systems during testing.

These episodes showed how hard it is in practice to control the unpredictable behavior of increasingly autonomous AI agents, even in test environments built by the companies developing the models themselves. OpenAI openly acknowledges that these cases are what pushed it to reconsider some of the very regulations it had previously criticized.

California as a Testing Ground

OpenAI chose to back four California state bills, seen as a seedbed for future federal legislation. SB 813 and AB 1405 would establish a framework for independent evaluation of AI systems and standards for auditors of those systems. AB 1864 addresses safeguards against using AI to create biological threats, while SB 1119 introduces protections for children and teenagers using chatbots. Governor Gavin Newsom signed two of the four bills on September 10, 2026.

This mechanism, in which states set de facto national standards before Congress does, is sometimes called reverse federalism. Critics point out, however, that a regulatory floor a large company can clear easily may force a smaller player to build an entire new compliance department.

What It Means for Poland and Europe

In broad shape, OpenAI's proposal resembles the EU's AI Act, though Brussels had already opted to soften some provisions after criticism from startups and small and medium-sized businesses. For Polish companies developing or deploying AI systems, this suggests the US market could, in the coming months, start moving closer to the European regulatory model built on risk assessment and mandatory testing, rather than the largely voluntary industry-commitment approach it has followed so far.

OpenAI said it would continue supporting state-level legislation until Congress acts at the federal level, and called on lawmakers to pass rules before the current congressional session ends. The company also said credible national standards are a precondition for the United States to lead in building compatible international standards.