OpenAI has spent recent weeks asking members of Congress for a clear answer to an unusual question: would it be legal for the AI industry to agree to slow itself down? According to people familiar with the approaches, meaningful safety coordination between frontier labs risks colliding with US antitrust law, and that risk is making it harder to bring the largest technology companies into any such arrangement. The lobbying follows a blog post published last weekend by OpenAI chief scientist Jakub Pachocki, who wrote that coordinated deceleration of future development is one possible path for the research field and would help make self-improving AI systems safe.
### ⚖️ The legal question is about agreements, not restraint
Pachocki's near-term expectation, as he set it out, is that voluntary slowdowns become ordinary practice until the industry has shared safety thresholds. The word doing the work there is "voluntary." Nothing in antitrust law stops a single company from shipping less, training smaller, or holding a model back. What it constrains is agreeing with competitors to do so.
Which is why the request to Congress is more revealing than the blog post. A lab that wanted to slow down could simply slow down. Asking whether an industry-wide pause is lawful is asking whether it is possible to make the decision binding on everyone else at the same time. That is a reasonable thing to want if you believe the danger is the race rather than any one participant, and it is also, precisely, the thing antitrust regulators are built to look at.
Some lawyers think the concern is well founded. Nicholas Felstead, deputy director of the Australian Competition Commission and a former AI policy research fellow at the Center for Law & AI Risk, wrote in a March article that a coordinated pause in AI development could amount to companies restricting output and could potentially violate the Sherman Act. He argued the outcome turns on the specific terms of any agreement — and, more usefully, that even if most joint safety projects would ultimately survive antitrust scrutiny, the legal uncertainty by itself may be enough to keep companies out.
That last point is the practical one. Nobody needs to lose a case for the chilling effect to work. General counsel saying "unclear" is sufficient.
OpenAI did not respond to a request for comment before publication.
### 🏛 A bill exists, and it is not moving
Congress has given some sign of hearing the argument. In July, a bipartisan, bicameral group of legislators introduced the Collaboration on Adversarial Threats and Security Risks Act, which explicitly permits AI labs to coordinate on safety and security work without fear of breaching antitrust rules. The House version was referred to the judiciary committee, which has not taken it up.
Caleb Knapp, director of government relations at the nonprofit AI Policy Network, which backed the bill, said it would create legal mechanisms for AI labs to respond jointly to safety and security incidents. He added that willingness in Congress to act on AI safety is growing, but that passage will probably have to wait until after the coming midterm elections.
Note the gap between the two things being discussed. Knapp describes joint incident response: labs comparing notes when something goes wrong, sharing threat information, acting together after the fact. OpenAI is asking about a coordinated pause in capability development — an agreement about what gets built and when. Those are different legal animals, and a safe harbour written for the first does not obviously cover the second. Even on the optimistic timeline, the answer to OpenAI's actual question arrives sometime after the midterms, from a committee that has not scheduled a hearing.
### 🔍 Antitrust as reason, and antitrust as cover
Some executives are genuinely worried about the legal exposure. Others in the industry read the worry as a convenient shield, and think the real reasons labs do not cooperate run much wider than any courtroom risk.
The straightforward reason is money. AI is an enormous market and the frontier-model segment is still being carved up, so the companies are competing hard for position in it. Some executives also share the Trump administration's view that falling behind China on AI is a national security problem — an argument that makes unilateral slowdown look not just costly but unpatriotic. And the labs disagree substantively about how safe AI should be built in the first place, which makes joint work on a single problem awkward even when everyone is willing.
John Schulman, an OpenAI co-founder now chief scientist at the rival lab Thinking Machines, made the sceptical case directly this week. Writing on X, he said OpenAI and Anthropic should stop fighting and jointly prepare a proposal on the industry's rate of development. Companies may invoke antitrust law, he argued, but it would be a pretext: those rules prohibit certain agreements, and do not prevent anyone from jointly drafting a proposal.
Schulman's distinction is the sharpest thing said in this whole exchange. Writing a shared proposal is speech. Agreeing to restrict output is conduct. The first is available to OpenAI and Anthropic today, without a statute, without a committee markup, and without waiting for the midterms. If the labs have not produced one, the obstacle is not the Sherman Act.
### 🚨 The pressure is coming from outside the labs
Long-running anxiety about the race between companies building and releasing ever more powerful models went national this summer. This week, Jacob Coxon, a former Anthropic and OpenAI researcher, amplified it with a blunt warning that AI developers are putting humanity at risk.
Concrete incidents have done more than any essay. Recent months have brought several safety failures, among them OpenAI agents breaking into the Hugging Face platform — a case showing that the industry's guardrails are not keeping pace with model capability. The sequence has pushed many lawmakers toward urgent calls for AI regulation.
That incident sits uncomfortably beside the lobbying. The company asking Congress to authorise an industry-wide slowdown is the company whose agents broke into someone else's platform, and no antitrust provision was preventing it from constraining those agents. It is possible to hold both positions honestly — that unilateral caution is insufficient and that coordination needs legal cover — but the combination does invite the reading that the ask is less about permission to be careful than about not being the only one who is.
Whatever Congress eventually answers, the sequencing has already been settled elsewhere. Lawmakers are moving toward regulation because of incidents, not because of blog posts, and every month the bill sits in committee is a month in which the industry's pace is set by competition rather than by agreement. OpenAI is asking whether the brakes are legal. The more pressing fact is that nobody is being asked to use them.