OpenAI Asks Congress: Is a Coordinated AI Slowdown Even Legal?
OpenAI has asked Congress whether coordinating an industry-wide AI slowdown would breach antitrust law, as a bipartisan bill tries to clear the legal path.
OpenAI has quietly approached members of Congress in recent weeks asking for legal guidance on whether orchestrating an industry-wide slowdown on frontier AI development would run afoul of US antitrust law, according to people close to the company who spoke to WIRED. The question follows a public blog post by OpenAI chief scientist Jakub Pachocki calling for coordinated slowdowns as a safety measure. A bipartisan bill introduced in July could create a legal path, but it has not yet been taken up by committee.
What happened
| Detail | Fact |
|---|---|
| Who asked Congress | OpenAI, in recent weeks |
| Topic of the inquiry | Whether coordinating an industry-wide AI slowdown is legal under US antitrust law |
| OpenAI chief scientist | Jakub Pachocki, published a blog post last weekend |
| Pachocki’s position | Voluntary slowdowns should become “commonplace” until shared safety standards are established |
| Relevant law cited | The Sherman Antitrust Act (the main US law prohibiting anti-competitive agreements) |
| Bill introduced | Collaboration on Adversarial Threats and Security Risks Act, July, bipartisan and bicameral |
| Bill status | House version referred to the Judiciary Committee; not yet taken up |
OpenAI has been asking lawmakers directly whether safety-focused coordination with rival AI labs could be treated as an illegal restriction of output. According to people familiar with the discussions, the concern is that any substantive agreement between companies on development pace could violate the Sherman Antitrust Act, which prohibits companies from coordinating in ways that harm competition.
The backdrop is a blog post published last weekend by Jakub Pachocki, OpenAI’s chief scientist. Pachocki argued that coordinating to slow down future AI development is a necessary path to ensuring that self-improving AI systems remain safe. He expects voluntary slowdowns to become routine before the industry settles on shared safety benchmarks.
What the legal risk actually looks like
Nicholas Felstead, assistant director of the Australian Competition and Consumer Commission and a former AI policy fellow at the Center for Law and AI Risk, wrote in a March article that a coordinated development pause could amount to companies restricting output, a classic antitrust violation. His conclusion was nuanced: the outcome depends “entirely on the precise details of any agreement,” but even if most safety collaborations would survive scrutiny, legal uncertainty alone “can act as a powerful deterrent.”
The bipartisan Collaboration on Adversarial Threats and Security Risks Act, introduced in July, is designed to cut through that uncertainty. According to Caleb Knapp, director of government affairs at the nonprofit AI Policy Network (which endorsed the bill), it would create explicit legal channels for AI labs to coordinate on safety and security without antitrust risk. Knapp told WIRED that Congress has a “growing appetite to get something done” on AI safety, but that passing anything into law may have to wait until after the upcoming midterm elections.
Is antitrust really the obstacle?
Not everyone accepts the legal risk as the main barrier. John Schulman, an OpenAI cofounder who is now chief scientist at the rival lab Thinking Machines, wrote on X this week that the antitrust argument is “fake.” His reasoning: antitrust law prohibits certain agreements, not the act of jointly developing a proposal. He called on OpenAI and Anthropic to stop feuding and work on a pacing proposal together as a first step.
A separate camp of AI executives points to more concrete tensions. These include fierce commercial competition over a nascent market for frontier models, disagreements between companies about what safe AI development even looks like, and the view (shared by some in the Trump administration) that falling behind China in AI capability is a national security risk. These factors may matter more than legal liability.
The urgency in Washington has been building. This summer, a string of security incidents, including OpenAI’s AI agents reportedly hacking Hugging Face, drew attention to gaps between rising model capability and the safeguards meant to contain it. Former Anthropic and OpenAI researcher Jacob Coxon added to the pressure this week with a public warning that AI developers are putting humanity at risk. As we covered in our piece on Coxon’s departure from Anthropic, the mood among safety researchers has shifted noticeably.
Our take
The antitrust question is real, but Schulman’s point is worth sitting with: there is a big difference between an agreement to slow down and a jointly written proposal about how slowing down could work. The first is legally fraught; the second is arguably just lobbying. OpenAI going to Congress first, rather than picking up the phone to Anthropic, tells you something about the internal politics here.
For businesses that depend on frontier AI tools, the practical implication is a period of regulatory uncertainty. If a slowdown bill passes, expect shorter capability jumps between model releases and longer gaps between major product updates from labs. That is not necessarily bad. More time for safety testing could mean fewer abrupt API changes and fewer situations where a model you built on suddenly behaves differently after a surprise update.
If you are building products on top of AI APIs right now, this is a good moment to think about abstraction layers. Our AI integration work consistently shows that businesses that avoid hard-coding to a single model version survive provider changes far better than those that do not.
What to do about it
- Watch the Judiciary Committee for movement on the Collaboration on Adversarial Threats and Security Risks Act. Its passage (or failure) sets the legal boundary for what labs can coordinate on.
- Audit any AI tools your business relies on. Identify which depend on specific model versions and which are model-agnostic.
- Build or request abstraction layers in your AI integrations so you can swap providers or model versions without rewriting core logic.
- Track the midterm elections. Knapp’s comment suggests major AI legislation is unlikely before then, so plan your AI roadmap with at least a 12-month regulatory horizon in mind.
The companies that will be least disrupted by an AI slowdown are the ones that already treat model providers as interchangeable infrastructure.
Frequently asked questions
Could coordinating an AI slowdown violate US antitrust law?
Possibly. Legal scholars warn that a coordinated pause in AI development could be treated as companies agreeing to restrict output, which may violate the Sherman Antitrust Act. The outcome depends heavily on the specific terms of any agreement, but even legal uncertainty can deter labs from cooperating.
What is the Collaboration on Adversarial Threats and Security Risks Act?
It is a bipartisan, bicameral bill introduced in July that would explicitly allow AI labs to coordinate on safety and security work without risking antitrust violations. The House version has been referred to the Judiciary Committee but has not yet been taken up.
What did OpenAI's chief scientist say about slowing down AI?
Jakub Pachocki, OpenAI's chief scientist, published a blog post arguing that coordinating to slow down future AI development is key to ensuring self-improving AI systems remain safe. He expects voluntary slowdowns to become commonplace until shared safety standards are established.
Why did John Schulman say the antitrust concern is 'fake'?
Schulman, an OpenAI cofounder now at Thinking Machines, argued on X that antitrust law prohibits certain agreements, not the act of jointly developing a proposal. He called on OpenAI and Anthropic to stop feuding and collaborate on a pacing proposal together.
