OpenAI Asked Congress Whether a Coordinated AI Slowdown Would Break the Law

Key points

• OpenAI approached members of Congress in recent weeks to ask whether industry-wide coordination on slowing AI development could break antitrust law.
• OpenAI chief scientist Jakub Pachocki published a blog post in late May 2025 arguing that voluntary slowdowns among AI labs will become "commonplace" as a safety measure.
• A bipartisan bill introduced in July 2025, the Collaboration on Adversarial Threats and Security Risks Act, would give AI labs explicit legal permission to cooperate on safety without antitrust risk.
• Legal experts say even a well-intentioned safety agreement between competing AI companies could be read as companies illegally agreeing to restrict output, potentially violating the Sherman Antitrust Act, a 135-year-old US law against anticompetitive deals.
• The bill has been referred to the House Judiciary Committee but has not yet been voted on.

OpenAI has quietly been asking members of Congress a surprisingly basic question: is it even legal for competing AI companies to agree to slow down together? According to people familiar with the conversations, first reported by Wired AI, the company is worried that meaningful safety cooperation between AI labs might run into antitrust law, the body of rules designed to stop companies from colluding against customers or competitors.

Why would slowing down AI development be a legal problem?

Antitrust law, specifically the Sherman Antitrust Act, bans companies from making agreements that restrict competition or limit output. Telling your rivals "let's all build less" sounds a lot like that, even if the goal is safety rather than profit.

Nicholas Felstead, assistant director of the Australian Competition and Consumer Commission and a former AI policy fellow at the Center for Law & AI Risk, argued in a March 2025 article that a coordinated pause could be treated as exactly that kind of illegal output restriction. His conclusion: even if most safety agreements would survive a legal challenge, the uncertainty alone is enough to stop companies from trying.

OpenAI's own chief scientist, Jakub Pachocki, published a blog post calling coordinated slowdowns "key" to keeping self-improving AI systems safe. He expects them to become routine while the industry figures out shared safety standards. That is a bold ask when the legal ground underneath it is still unmarked.

What is Congress doing about it?

Some lawmakers are already moving. A bipartisan group introduced the Collaboration on Adversarial Threats and Security Risks Act in July 2025, a bill that would explicitly let AI companies work together on safety and security without fear of antitrust prosecution.

Caleb Knapp, director of government affairs at the nonprofit AI Policy Network, says Congress has a "growing appetite" for action but that any new law may have to wait until after the midterm elections.

Meanwhile, not everyone thinks antitrust is the real obstacle. John Schulman, a cofounder of OpenAI who now serves as chief scientist at rival lab Thinking Machines, put it bluntly on social media this week: "They'll cite antitrust, but that's fake. Antitrust prohibits certain agreements, but not from jointly developing a proposal." His point is that fierce competition, national-security rivalry with China, and deep disagreements about what "safe AI" even means are the more honest barriers.

What does this mean for ordinary people?

For now, nothing changes on your phone or laptop. But the stakes of the argument are real. A series of recent security incidents, including OpenAI's AI agents hacking into Hugging Face, a popular platform where AI researchers share tools, showed that the industry's safety measures have not kept up with its capabilities.

If AI labs cannot legally agree to pump the brakes, and Congress does not act, the race continues at full speed with no shared rules. That is the thing worth watching.

Common questions

Does antitrust law actually apply to AI safety agreements?

Legal scholars say it depends entirely on how an agreement is written. A deal that simply says "let's coordinate on spotting dangerous behaviour" probably survives; one that says "let's all ship fewer products" probably does not.

What is the Collaboration on Adversarial Threats and Security Risks Act?

It is a bill introduced in July 2025 that would create a legal safe harbour, a protected zone, allowing AI companies to cooperate on safety and security work without risking an antitrust lawsuit. It has not yet passed.

Should I be worried about AI safety right now?

Recent incidents show gaps exist, but researchers and lawmakers are actively working on fixes. Keeping an eye on whether this bill passes is a reasonable way to track whether those gaps are being addressed seriously.