Freezing frontier AI: the lawsuit against OpenAI, Anthropic, Google, Meta and SpaceXSI, and the EU AI Act

In one sentence: On 12 November 2026, four US nonprofits asked the Superior Court of California for the County of San Francisco to bar Anthropic, Google, Meta, OpenAI and SpaceXSI from developing or deploying AI more capable than their current systems; they rely solely on California's unfair competition law, whereas the EU AI Act regulates how large models are placed on the market, not how they are trained.

The press releases speak of a "freeze" on frontier AI. The complaint itself, read in full by Le Latent, is more precise: it says who is suing, on which statute, and what the court should order. It also shows where the companies are likely to push back. And it reads differently from Europe, where the same risks are already named in a text that three of the five companies have signed. For our wider coverage, see the Society section.

Who is suing whom, and in which court

The complaint, dated 12 November 2026, was filed in the Superior Court of the State of California for the County of San Francisco, a state court rather than a federal one. The four plaintiffs are Irreplaceable Education Fund, Public Citizen, Heat Initiative and Future Tech Insights, which operates as The Alliance for Secure AI. Their counsel include Matthew Platkin, a former Attorney General of New Jersey, and the firms Gerstein Harrow and Legal Advocates for Safe Science and Technology.

Seven entities are named: Anthropic PBC, Google LLC, Meta Platforms, OpenAI Group PBC, OpenAI Foundation, SpaceX Super Intelligence LLC and SpaceX Super Intelligence Corp. According to the complaint, the last two are Nevada entities that develop Grok, the model originally built by xAI. OpenAI is sued through both its for-profit company and the foundation that appoints its board.

The incidents behind the request are those the complaint attributes to the companies' AI agents: the intrusion into Hugging Face's infrastructure disclosed on 16 July 2026, which the complaint links to OpenAI models under internal testing, three incidents Anthropic disclosed on 30 July 2026, and a false tip sent to the Philadelphia police. These are allegations, sourced to company posts and press reports, that the court has not examined.

What the nonprofits are asking for

The complaint ends with six requests. The first four are "immediate" injunctions; none seeks money for the plaintiffs, apart from attorneys' fees.

Source: complaint, Prayer for Relief, read on 14 November 2026.

Request B targets what the complaint calls recursive self-improvement. To support it, the complaint cites Anthropic's IPO filing and the METR report published on 5 November 2026, which we have analysed (in French): the safety threshold in Anthropic's amended S-1 and what METR's report says about AI-driven research at Anthropic.

Request A sets no measure of "capability": no compute threshold, no test score. If the injunction were granted, it would be for the court to decide which model exceeds a company's most capable one as of 12 November 2026.

The legal basis: an unfair competition statute

The complaint contains a single cause of action: the Unfair Competition Law, section 17200 and following of the California Business and Professions Code. The statute covers any "unlawful, unfair or fraudulent business act" and allows an injunction against anyone who engages, or "proposes to engage", in one. The complaint relies on the first two prongs.

The "unlawful" prong borrows another law: here, public nuisance, which the California Penal Code defines as anything "injurious to health" or obstructing the free use of property so as to affect a considerable number of people. The plaintiffs argue that building systems that pose catastrophic risks, or that attack third-party computer systems, creates such a nuisance.

The "unfair" prong invokes three California public policies, including SB 53, the Transparency in Frontier Artificial Intelligence Act, in force since 1 January 2026. The complaint does not sue the companies for breaching SB 53; it cites the law as evidence of legislative intent. Section 22757.15 reserves SB 53's civil penalty, up to one million dollars per violation, to an action "brought only by the Attorney General".

What could stop the case

The published documents contain no response from the companies and no hearing date. Several pressure points can still be read in the text itself.

Standing. Since Proposition 64, adopted by ballot initiative on 2 November 2004, section 17204 limits private suits to a person "who has suffered injury in fact and has lost money or property as a result of the unfair competition". That is why the complaint spends several pages on the resources each nonprofit says it had to divert, such as a student walkout brought forward to 20 October 2026, and states that these choices predate the lawsuit. The court will decide whether that is enough.

Jurisdiction over SpaceXSI. The complaint asserts general jurisdiction over Anthropic, Google, Meta and OpenAI, which are domiciled in California, but not over the two Nevada SpaceX entities: for those, it points to Grok research carried out in Palo Alto.

The scale of the remedy. A state court is being asked to halt a whole technology at five groups on the basis of a competition statute. The plaintiffs present this as applying existing law and cite rulings from other sectors, including a Long Beach refinery case; the complaint cites no ruling that has ordered such a halt for AI.

The sources we read report no reaction from the companies named.

The EU view: no freeze, but a market withdrawal

A European court asked by consumer groups could not grant request A on the basis of the AI Act. Article 2(8) of Regulation (EU) 2024/1689 states that the Regulation "does not apply to any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service". The AI Act governs models sold or made available in the EU, not laboratories.

It does provide something the California case lacks: a regulator. For general-purpose AI models with systemic risk, presumed above 10^25 floating-point operations of training compute (Article 51), the Commission can request documentation, evaluate the model and, "where necessary and appropriate", request the provider to "restrict the making available on the market, withdraw or recall the model" (Articles 91 to 93). Article 52 also requires the provider to notify the Commission within two weeks of knowing a model will cross the threshold, which can happen before training ends. On 14 November 2026, we applied that threshold to Mistral Large 4 on the basis of its technical report (in French).

Fines can reach 3% of worldwide annual turnover or 15 million euros, whichever is higher (Article 101). Provider obligations have applied since 2 August 2025, and fines and the Commission's enforcement powers since 2 August 2026, under Article 113; the 2026 omnibus regulation did not change either date. The full schedule is in our EU AI Act timeline after the 2026 delay.

What about civil society? Article 110 of the AI Act adds the Regulation to Annex I of the Representative Actions Directive (EU) 2020/1828, and the omnibus regulation made that change applicable from 27 July 2026. "Qualified entities" designated by each Member State can therefore seek injunctions against AI Act infringements that harm the collective interests of consumers, without having to prove actual loss suffered by individual consumers (Article 8). But there has to be an infringement: a compliant model cannot be stopped this way.

Sources: complaint, California Business and Professions Code, s. 17204, Regulation (EU) 2024/1689, Directive (EU) 2020/1828, checked on 14 November 2026.

The EU code already commits signatories not to proceed

The closest match between the complaint and EU law is in the Safety and Security chapter of the Code of Practice, published on 10 July 2025. Its list of specified systemic risks almost mirrors the nonprofits' grievances: chemical, biological, radiological and nuclear risks, cyber offence, and "loss of control", which covers self-improvement and "autonomously creating or improving AI models or AI systems".

Above all, its Measure 4.2 says signatories "will only proceed with the development, the making available on the market, and/or the use of the model, if the systemic risks stemming from the model are determined to be acceptable", and otherwise will not make it available, or will restrict, withdraw or recall it. This is a voluntary commitment, with acceptance criteria set by each company, not a court order. According to the Commission's list, Anthropic, Google and OpenAI have signed the code, xAI only its Safety and Security chapter, and Meta is not listed; a provider of a systemic-risk model that does not sign must demonstrate "alternative adequate means of compliance" (Article 55).

The idea of a shutdown mechanism also appears in the Paris Call for AI safety of 9 November 2026, without any binding force, and in OpenAI's Control Room kill switch.

What it means for you

If you use ChatGPT, Claude, Gemini, Grok or Meta's assistants in Europe, nothing changes: the published documents mention no injunction, and an order from a California court would target the development of new models, not existing services.

If you are a lawyer, a data protection officer or in a compliance team, the complaint is above all a compilation: its 279 paragraphs gather, with references, the incidents and public statements of the five groups since summer 2026 (complaint). In an EU vendor audit, the useful questions are different: has the provider signed the code, what acceptance criteria has it published, has it notified its model as one with systemic risk?

If you follow the policy debate, note the difference in method: in California, nonprofits are asking a state judge to act under a general statute; in the EU, a law specific to AI gives the Commission the power to have a model taken off the market.

What we checked, and what we do not know

We read the complaint as published by the plaintiffs' counsel and reproduced by Just Security, the press releases of Public Citizen and The Alliance for Secure AI, the summary by Just Security, the California code sections cited and the EU texts in the Official Journal.

What the documents do not say: the case number (the published copy leaves it blank), the date of a first hearing, and whether a separate motion for a preliminary injunction has been filed. Several of the plaintiffs' lawyers were still awaiting permission to appear in California. The incidents described remain allegations. This article describes legal texts; it is not legal advice.