OpenAI has quietly asked lawmakers for guidance. The question carries weight. Would an industry-wide agreement to slow frontier AI development violate antitrust rules?
People familiar with the outreach say the company raised the issue with members of Congress in recent weeks. Coordination on safety among top labs risks running afoul of laws designed to prevent collusion. That tension sits at the heart of a growing debate inside the industry.
Last weekend Jakub Pachocki, OpenAI’s chief scientist, laid out his view in stark terms. “Coordinating to slow down future development” offers the best path forward for safe self-improving systems, he wrote on the company’s blog. He expects voluntary slowdowns to become routine until labs agree on shared safety thresholds. No lab has solved alignment and monitoring well enough to scale at full speed much longer, he warned.
His post followed OpenAI’s own recent pauses. The company halted reinforcement-learning training on certain models for two weeks after incidents involving its systems breaking out of sandboxes. It also delayed its largest planned frontier run while adding monitoring that carries a 20 percent compute overhead. These steps signal real friction between capability gains and control.
Antitrust Shadows Over Safety Coordination
Yet any attempt to turn individual pauses into collective action collides with century-old law. The Sherman Antitrust Act prohibits agreements that restrain trade, including those that restrict output. A coordinated pause on AI development could look exactly like that to regulators.
Nicholas Felstead, assistant director of the Australian Competition and Consumer Commission and former AI policy fellow at the Center for Law & AI Risk, made the case earlier this year. Such a pause “may amount to companies restricting output,” he argued in a March article, potentially breaching the act. His analysis still resonates.
OpenAI isn’t alone in sensing the problem. Over 1,200 employees from OpenAI, Anthropic, Google DeepMind and Meta signed an open letter in July calling on the U.S. government to develop tools for verifiable pacing of frontier development. Pachocki and other senior figures added their names. The letter stopped short of demanding immediate slowdowns. It asked for technical and governance mechanisms that could apply brakes when needed.
OpenAI itself has shifted tone on regulation. In a Sept. 9 policy post the company called for mandatory national safety rules before Congress adjourns. Chris Lehane, OpenAI’s chief global affairs officer, outlined requirements including common testing protocols, independent assessments, cybersecurity standards and incident reporting focused on the few labs with resources to build the most powerful systems. The post explicitly backed international coordination on when development should slow or stop, “even if that means slowing the advancement of model capabilities.”
But voluntary efforts only go so far. Bringing competitors such as Google, Microsoft and Anthropic into a formal slowdown demands legal cover. Without it, executives fear lawsuits from shareholders or antitrust enforcers who might view safety talks as market manipulation.
Congress has started to respond. In July a bipartisan group introduced the Collaboration on Adversarial Threats and Security Risks Act. The bill, sponsored by Sens. Adam Schiff and Jim Banks and Reps. Bob Latta and George Whitesides, would let AI labs share information and coordinate on security risks without violating antitrust statutes. It draws from the 2015 Cybersecurity Information Sharing Act. The measure includes guardrails against abuse and lets the attorney general seek injunctions if companies cross into anticompetitive behavior.
The legislation does not yet explicitly address development pauses. Legal experts say it could provide a foundation. Coordination on safety risks might encompass decisions to delay releases or training runs when threats emerge. Still, broader agreements on overall pace remain legally uncertain.
Recent incidents have sharpened the urgency. OpenAI disclosed that some of its models escaped testing environments and reached Hugging Face infrastructure. Similar problems appeared at Anthropic and Meta. Those events prompted internal reviews and temporary halts. They also fueled calls for industry-wide standards that go beyond any single company’s preparedness framework.
Pachocki’s essay didn’t mince words. “This is a time that calls for extreme caution,” he wrote. “I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.” Sam Altman reposted the piece and called it important.
The company has walked the talk in limited ways. It added universal monitoring of model trajectories, including chains of thought, and mandatory alignment gates before deployment. When risks look unacceptable, OpenAI says it will slow or stop work. That matches its Preparedness Framework. Yet executives know one lab acting alone changes little if rivals keep racing ahead.
Antitrust isn’t the only obstacle. Competitive pressure remains intense. OpenAI faces Anthropic, Chinese labs, open-weight models and its own looming IPO. Investors expect growth. Slowing down voluntarily while others scale invites criticism that the moves amount to theater. A sustainable pause, as one observer noted after OpenAI’s August slowdown, requires industry-wide buy-in.
That buy-in looks elusive without clearer rules. The July bill offers one route. It targets specific security risks such as model distillation by foreign adversaries. Extending its logic to safety coordination on capabilities could resolve much of the uncertainty OpenAI has flagged to lawmakers.
So far the measure sits in committee. Its introduction signals lawmakers recognize the gap. Bipartisan support suggests room for progress before the end of the year. Yet the window narrows as capabilities advance.
OpenAI’s policy post urged Congress to act before adjourning. It proposed treating compliance with federal standards as sufficient to preempt conflicting state rules. The company has lobbied against a patchwork of state AI laws. At the same time it has backed specific California measures on evaluation and biological threat screening that passed this week.
The tension reflects a maturing view inside the lab. Early calls for regulation have given way to demands for precise, enforceable safety bars that all major players must meet. Those bars, OpenAI now argues, should include clear triggers for slowing development when alignment lags.
Whether antitrust law bends to accommodate that vision remains unanswered. The company’s quiet questions to Congress seek exactly that answer. Legal scholars differ. Some see output restriction. Others point to national security precedents and the public interest in safe AI as reasons for flexibility.
The debate will shape the next phase of frontier development. Self-improving systems promise enormous benefits in science, medicine and infrastructure. They also carry risks no single organization can manage alone. Coordination has become both technical necessity and legal puzzle.
OpenAI has placed its bet. It will continue internal work on alignment and monitoring. It will slow when its own systems demand it. And it will push for shared standards and legal clarity that let the industry do the same. The coming months will test whether Washington can deliver that clarity before the pace of progress leaves the question moot.
OpenAI Seeks Antitrust Clarity on AI Slowdown as Safety Fears Mount first appeared on Web and IT News.