OpenAI has a new legal problem, except this time it’s one they’re looking for.

The company recently asked California’s attorney general whether coordinating with competitors to slow AI development would violate antitrust law. It’s a strange question for a company to pose voluntarily, but it reflects a real tension in the AI industry right now. Some researchers think the technology is advancing too fast. Some think coordination is necessary to prevent catastrophic outcomes. And antitrust law wasn’t written with any of this in mind.

The basic issue is simple. When competitors agree to limit output or slow production, that’s usually illegal. It’s price-fixing’s cousin. The Sherman Act calls it restraint of trade, and courts don’t like it. But what if the companies argue they’re not trying to limit competition, they’re trying to prevent everyone from dying?

That’s not hyperbole. A growing number of AI researchers, including people inside major labs, believe advanced AI systems pose genuine extinction risk. A recent report suggested that concerns about recursive self-improvement and agentic swarms are “genuinely spooking people” at companies like OpenAI, Anthropic, and Google DeepMind. If you take that risk seriously, then asking competitors to pump the brakes starts to sound less like collusion and more like self-preservation.

But antitrust law doesn’t care much about your motivation. The relevant question is whether the agreement harms competition. If OpenAI, Google, and Anthropic sat down and agreed to cap their model releases at a certain capability threshold, that would look an awful lot like market division. It doesn’t matter if they’re doing it to save humanity. The law treats agreements between competitors as inherently suspect.

Why OpenAI is asking now

OpenAI isn’t asking this question in a vacuum. There’s already pressure building for some kind of industry-wide coordination on AI safety. Researchers have called for pauses. Policymakers have floated mandatory safety thresholds. And some inside the labs themselves worry that competitive pressure is pushing development faster than safety work can keep up.

The problem is that any coordination between the handful of companies at the frontier immediately raises antitrust red flags. There are maybe five companies in the world training models at the scale of GPT-5 or Gemini Ultra. If those five agreed to slow down, you’d have a concentrated market with horizontal agreements among direct competitors. That’s the kind of thing the FTC and DOJ usually file suit over.

OpenAI’s inquiry suggests they’re aware of this. By asking California’s AG for guidance, they’re essentially seeking a roadmap for how to coordinate on safety without getting sued. It’s unclear whether they’ll get one. Antitrust enforcers have historically been skeptical of efficiency justifications for coordination, and “we’re trying not to build Skynet” is a novel argument even by those standards.

The legal landscape

There are a few ways this could play out legally. One possibility is that AI companies argue for a kind of safety exception under the rule of reason, the framework courts use to evaluate whether agreements unreasonably restrain trade. The argument would be that while coordination limits competition in one sense, it prevents a much worse outcome, total market collapse (or human extinction, depending on how seriously you take the risk).

Courts have occasionally allowed coordination when it’s necessary to make a market function at all. Sports leagues, for example, get some leeway because competition requires teams to agree on rules. But those exceptions are narrow, and it’s hard to see how slowing AI development fits the same mold. The market works fine without coordination. It just might produce something dangerous.

Another route is legislation. Congress could pass a law explicitly allowing or requiring coordination on AI safety under certain conditions. That’s what happened with standard-setting organizations, which get limited antitrust immunity when they’re setting technical standards. But getting Congress to act requires consensus on what the risk actually is, and right now there’s no agreement on whether advanced AI is an extinction threat or an overhyped concern.

The third option is that companies just don’t coordinate, and we find out the hard way whether the risks were real. That’s the default outcome if antitrust law makes coordination too risky. Some researchers think that’s already happening, that competitive pressure is making it nearly impossible for companies to slow down even when their own safety teams are raising alarms.

What happens next

OpenAI’s inquiry to California probably won’t produce binding guidance. State AGs can offer opinions, but they don’t control federal antitrust enforcement, and they can’t immunize companies from private lawsuits. At best, OpenAI might get a sense of how regulators would view coordination. At worst, they’ve just put a target on their back by announcing they’re considering it.

The broader question is whether antitrust law can adapt to a situation where the usual logic doesn’t apply. The whole point of antitrust is to prevent companies from limiting competition to jack up prices or restrict output. But if the output is something that might pose catastrophic risk, limiting it starts to sound reasonable.

Courts have generally resisted these kinds of arguments. When tobacco companies tried to coordinate on safety standards, antitrust enforcers sued them anyway. When automakers wanted to jointly develop emissions technology, they had to get explicit permission. The bar for justifying competitor coordination is high, and “we’re worried about what we’re building” might not clear it.

That leaves the industry in a strange position. The companies building the most advanced AI systems are also the ones most worried about where it’s heading. But the law that’s supposed to keep them honest also makes it illegal for them to collectively slow down. Whether that’s a feature or a bug depends on whether you think the risk is real.