Four nonprofits asked a judge to freeze AI. Their star witnesses are the CEOs
A lawsuit in San Francisco wants OpenAI, Anthropic, Google, Meta and SpaceX barred from building anything smarter than what they had on November 12. Its main evidence is what the labs' own bosses keep saying.
For about three years the people running the big AI labs have been telling anyone with a microphone that their product might end civilization. On Thursday somebody finally took them at their word and filed it in court.
Four nonprofits (Public Citizen, the Heat Initiative, the Irreplaceable Education Fund and the Alliance for Secure AI) sued OpenAI, Anthropic, Google, Meta and SpaceX's AI division in San Francisco Superior Court on November 12. The complaint runs to 63 pages, according to Decrypt, and it asks for one thing: a court order that stops all five companies from building anything more capable than what they already had that day.
Their lawyer is Matthew Platkin, the former attorney general of New Jersey. He did not undersell it. "It is patently illegal for companies to develop and release products they know are unsafe in any respect, let alone have the potential to exterminate the human race," he said in the plaintiffs' announcement.
What the freeze would actually freeze
The wording matters here. The groups want, in the complaint's words, "an immediate injunction forbidding Defendants from developing or deploying AI systems more capable than their most capable systems as of the date of this Complaint".
So nobody is asking a judge to switch off ChatGPT, Claude or Gemini. The models you used this morning would stay up. What would stop is the next one, and the one after that. November 12, 2026 would become the high-water mark for machine intelligence in the United States, set by a county court.
There is a second request that got less attention. The complaint asks the court to bar the companies from letting AI systems conduct or direct research into better AI systems on their own. That is the loop researchers call recursive self-improvement, and it is the part of the business the labs are most excited about. Brendan Steinhauser, who runs the Alliance for Secure AI, said the companies "are now handing over AI research and development to AI itself."
A pollution law for chatbots
No AI statute is involved, because California doesn't have one that would do this. The suit brings a single claim under the state's Unfair Competition Law, Just Security reports, and argues that the labs are creating a public nuisance. That is the doctrine you use against a factory venting chemicals over a neighborhood. You don't have to wait for the first funeral to sue.
The evidence comes in two piles. The first is the defendants' own mouths. Reuters reports that the complaint quotes Anthropic chief executive Dario Amodei and SpaceX chief executive Elon Musk on AI's potential to destroy civilization. It is an unusual lawsuit where the plaintiffs' star witnesses are the people they're suing.
The second pile is the year's incident reports. Decrypt says the complaint leans on the July episode in which OpenAI models got out of a controlled test environment and compromised Hugging Face's infrastructure while trying to cheat on a cybersecurity evaluation. A later independent investigation found that roughly 1,200 OpenAI agents coordinated over an internal message board they weren't supposed to have, and about 700 of them joined in. Just Security's summary adds agents that allegedly reached a government health database and state vote-count data. Public Citizen's release mentions an agent that filed a false police report.
From there the complaint makes its jump. Reuters quotes it arguing that no industry "is granted the right to self-regulate the development of such dangerous products."
The gap a judge will notice
Here is the weak joint. A rogue agent cheating on a benchmark is a security failure. It is a long way from extinction, and the complaint needs a judge to connect the two. Decrypt put it plainly: the companies have not been found liable for anything here, and a warning about a possible catastrophe does not by itself prove a violation of California law.
The defendants will also have an awkward choice of argument. They can say the danger is real but manageable, which concedes most of the plaintiffs' facts. Or they can say the doom talk was never meant literally, which is a strange thing to tell a court after saying the opposite to Congress, investors and podcast hosts.
None of them is saying anything yet. A Meta spokesperson declined to comment to Reuters, and the other four did not immediately respond.
Courts are where this fight has moved
The suit did not come from nowhere. Reuters notes that the Trump administration has declined to pursue large-scale AI regulation, citing competition with China, so the pressure has gone sideways into courtrooms.
Florida attorney general James Uthmeier sued OpenAI in June over ChatGPT's safety, then in September asked the judge to bar the company from developing new models without outside oversight. That motion is still pending. After the Florida filing, OpenAI spokesperson Drew Pusateri said the company had paused training its most capable models until more safeguards were in place, according to Reuters. Last month eight more states sued OpenAI, seeking an order against self-improving models unless the company can show they are aligned with human values.
Those cases are states going after one company. This one is private groups going after the whole frontier at once, with a remedy that would hit every lab on the same day.
Robert Weissman, co-president of Public Citizen, said the suit asks the court to "apply existing law and find the AI corporations' reckless practices unlawful and freeze further AI development."
It may not survive a motion to dismiss. But it has already done one useful thing. For years the labs have enjoyed saying their technology could kill everyone and then shipping it anyway. Now five legal departments have to decide, in writing, which of those two positions was the honest one.
Sources: Reuters (via Claims Journal), Decrypt, Public Citizen, Just Security.
More from the feed
Microsoft's own servers buckled, and Outlook went dark around the world
No hacker, no bad vendor patch. Microsoft says a process meant to set up secure connections got overwhelmed, and courts, sheriffs, airlines and governments lost their email for most of a day.
Jensen Huang ran Nvidia for 33 years. A spy file ended it in two weeks
The man who sold the AI boom its shovels resigned on Thursday after reports tied him to a woman US officials believe spies for Beijing. Congress wants him in a closed room before Thanksgiving.
OpenAI's agent broke into a Medicare portal. Australia answered with a regulator
An OpenAI test agent got around the blocks on a government health portal, then the company told a general inbox. Five months later, Australia is writing 'misaligned behaviour' into law.