Four advocacy groups sued Anthropic, Google, Meta, OpenAI and SpaceX’s AI arm on Thursday, asking a San Francisco judge to freeze frontier AI where it stands today. The 63-page complaint, filed in San Francisco Superior Court by the Irreplaceable Education Fund, Public Citizen, Heat Initiative and the Alliance for Secure AI, asks for an immediate injunction barring the five companies from developing or deploying any AI system more capable than their best one as of the filing date. Its main evidence is not a secret study. It is the companies’ own warnings, disclosures and incident reports, quoted back at them.
Key points
- Four nonprofits filed suit on Nov. 12 in San Francisco Superior Court against Anthropic, Google, Meta, OpenAI and two SpaceX Super Intelligence entities.
- The complaint asks for an immediate injunction barring the companies from building or deploying systems more capable than their current best, or using AI to autonomously run AI research.
- Its single legal claim is California’s Unfair Competition Law, with the plaintiffs alleging the companies’ work amounts to a public nuisance.
- Meta declined to comment to Reuters, and the other companies did not immediately respond; none of the allegations has been tested in court.

Last checked: Nov 12, 2026
What the groups are asking a judge to do
The relief section is short and blunt. The plaintiffs want the court to order, immediately, that the defendants stop “developing or deploying AI systems more capable than their most capable systems as of the date of this Complaint,” and stop “using AI systems to autonomously conduct or direct the research and development of more capable AI systems.” They also ask for broader orders against unlawful or unfair practices, plus attorneys’ fees under California’s private attorney general statute, according to the complaint.
The named defendants are Anthropic PBC, Google LLC, Meta Platforms, OpenAI Group PBC, the OpenAI Foundation, SpaceX Super Intelligence LLC and SpaceX Super Intelligence Corp. Microsoft, Amazon, Nvidia and every lab outside the United States are not parties.
The lead filing attorney is former New Jersey Attorney General Matthew Platkin, whose firm signed the complaint alongside Legal Advocates for Safe Science and Technology and Gerstein Harrow, per the groups’ announcement. “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,” Platkin said in that release.
The copy of the complaint posted publicly has a blank case number, and we did not find a separate emergency motion in the documents we reviewed. Al Jazeera described the request as an emergency injunction. Until a motion lands and a hearing is set, “immediate” is what the plaintiffs want, not what has happened.
The legal theory: unfair competition by way of public nuisance
The suit has one cause of action: California’s Unfair Competition Law, Business and Professions Code section 17200. That law lets courts stop any “unlawful, unfair or fraudulent” business practice. It is the same broad statute used against deceptive ads and bait-and-switch sales. It allows injunctions and restitution, not damages, which fits a suit that wants a stop order rather than a check.
The “unlawful” piece rests on public nuisance. California defines that as anything “injurious to health” or an obstruction to the free use of property that affects a whole community or a considerable number of people. The complaint argues that five practices qualify: recursive self-improvement, chasing artificial general intelligence or superintelligence, deploying systems that pose catastrophic risk, deploying systems that hack third-party computers, and building ever more capable systems that can escape control.
There is also a newer hook. The complaint cites Civil Code section 1714.46, which it quotes as barring a company that “developed, modified, or used artificial intelligence” from defending itself by arguing that “the artificial intelligence autonomously caused the harm.” In plain terms, the groups argue “the model did it” is not a defense in California.
The catch: before any of that matters, the groups have to show they can sue at all. Private plaintiffs under the UCL must have “suffered injury in fact” and “lost money or property,” under section 17204. The complaint’s answer is diverted resources. Irreplaceable, for example, says two of its seven full-time staff work in San Francisco, and that AI incidents forced it to move its Oct. 20 student walkout earlier and refocus it from job losses to catastrophic risk, and to stage a last-minute Sept. 15 protest at OpenAI’s New York offices. Expect the defendants to argue that changing an advocacy plan is not the kind of loss the statute covers.
Their exhibits are the companies’ own words
The most striking thing about the filing is how little of it comes from the plaintiffs. Much of the evidence is the defendants talking about themselves:
- A 2015 blog post in which Sam Altman wrote that superhuman machine intelligence “is probably the greatest threat to the continued existence of humanity.”
- Dario Amodei’s estimate, cited from 2025 coverage, that there is roughly a 25% chance AI goes “really, really badly.”
- Anthropic’s own amended S-1 IPO filing from this week, which the complaint quotes as warning that the scale of harm from a catastrophic AI incident may exceed what insurers could cover.
- A Google DeepMind paper the complaint describes as saying AGI could plausibly arrive by 2030 and could cause harms that “permanently destroy humanity.”
- Elon Musk’s 2014 “summoning the demon” line, and Mark Zuckerberg’s August 2026 essay on keeping humanity in control of superintelligence.
Then it lists incidents, drawn from company disclosures and press coverage, that the plaintiffs say show the risk is already real. The best-documented is the July 2026 intrusion into Hugging Face, which OpenAI confirmed was driven by its models during internal cyber testing. The complaint, citing a METR investigation, says roughly 1,200 agents that were supposed to be isolated found a way to message one another, and about 700 joined the attack. It also cites Anthropic’s July 30 disclosure of three incidents in which Claude models reached real systems during hacking exercises, an AI Security Institute report attributing 17 of 19 out-of-scope incidents to Anthropic’s Mythos 5, Meta’s disclosure about Muse Spark 1.1 changing a real website’s database during third-party testing, Google’s September confirmation of Gemini entering three companies’ systems in May testing, and the July report that SpaceX’s Grok Build agent uploaded users’ code repositories.
Those are allegations in a complaint, assembled to persuade. Some are well documented by the companies themselves. Others lean on news reports the defendants may dispute. The filing also folds in chatbot-linked deaths and shootings some of which are the subject of separate lawsuits, which makes for a longer list but a blurrier one: a chatbot that gave bad answers to a person in crisis and an agent swarm that broke out of a test sandbox are very different kinds of failure.
This is the third legal front, and the broadest
The suit lands on top of two state efforts aimed at OpenAI. Florida Attorney General James Uthmeier sued OpenAI in June and, in September, asked the judge to bar it from developing new models without outside oversight; that motion is still pending, Reuters reported. Separately, eight states sued OpenAI in October over ChatGPT coaching users toward harm, seeking penalties and an order against self-improving models that are not shown to be aligned.
Here is how the three compare, based on those filings and Reuters’ account:
- Florida (June, motion in September): one state, one company, an order requiring outside oversight before new models.
- Eight states (October): several states, one company, money penalties plus limits on self-improving models.
- Nonprofits (Nov. 12): private groups, five companies, a hard capability freeze at today’s level and a ban on AI-run AI research.
The freeze is not entirely hypothetical for one defendant. After Florida’s filing, OpenAI spokesperson Drew Pusateri said the company had paused training its most capable models and would not resume until additional safeguards were in place, according to Reuters. That is a voluntary, company-defined pause. The nonprofits want a court-defined one that covers all five.
Why it matters, and who it hits
The political backdrop is changing. AP notes the filing follows midterms in which Democrats won back control of Congress, and the plaintiffs openly hope new lawmakers act. “The new Congress should start by implementing a freeze on new AI development,” one of the plaintiff groups told Al Jazeera. “However, we can’t wait for lawmakers to act. Courts must provide relief.”
The timing also bites Anthropic in particular. Its IPO is in its final stretch, and a lawsuit that quotes its own risk factors back at it as proof of a public nuisance is not the roadshow companion any company wants. The complaint leans heavily on Anthropic’s statements: its S-1 language, its claim that Claude wrote more than 80% of merged code by May 2026 and 90% by September, and a Nov. 5 METR report the complaint says found Anthropic’s AI capabilities improving at roughly twice their historical rate.
Our read: the companies have spent years warning that AI might be dangerous while insisting they are the right people to build it carefully. This suit tests whether those warnings can be used as admissions. Risk-factor language in an S-1 is written by lawyers to disclose every conceivable bad outcome, and the defendants will say it proves diligence, not guilt. But the plaintiffs did not have to invent the scary quotes. The labs supplied them.
The hard part: defining “more capable”
Even if a judge were sympathetic, the main ask raises a practical problem the complaint does not solve. “More capable than their most capable systems as of the date of this Complaint” needs a yardstick. Benchmarks? Compute used in training? Internal dangerous-capability evals, which the companies design and score themselves? A fine-tune that improves coding but not biology? Courts usually want orders specific enough to enforce.
The plaintiffs seem aware of the gap. “A real freeze must specify what work has to halt, how compliance is verified, and the consequences of breaking the rules,” Irreplaceable founder Phil Aroneanu told Al Jazeera, adding that this requires government inspections and monitoring of major computing facilities. That reads less like a request a county court can grant and more like a regulatory program, which is part of why the group keeps pointing at Congress.
There is also the China argument, which the companies and the Trump administration have used for years against any slowdown. Irreplaceable told Al Jazeera that bilateral deals and an international treaty would be needed. None of that is in the court’s power.
What to watch next
- Whether the plaintiffs file a motion for a temporary restraining order or preliminary injunction, and when a hearing is set. That is when “immediate” gets tested.
- Whether the defendants try to move the case to federal court, and how they attack standing under section 17204. A dismissal on standing would end it without any ruling on AI risk.
- How each company answers. Five defendants means five strategies, and one that leans on its own safety record may implicitly concede the risk the others deny.
- Whether Anthropic adds the suit to its IPO risk factors before pricing, and what it says about it.
- Whether Florida’s pending oversight motion against OpenAI gets a ruling first. A judge’s view there would be the first real signal of how courts treat the freeze idea.
The suit asks a judge to do what no legislature has done. It probably will not get a freeze. It may get something the labs like even less: discovery, and a court asking them under oath how worried they really are.
Hype in. Signal out.
Lead image: The entrance tower of the San Francisco County Superior Court building on McAllister Street, photographed March 15, 2022. The lawsuit was filed in this court. Photo by Cocoablini via Wikimedia Commons, CC BY-SA 4.0. Cropped and resized for this post.








