A coalition of nonprofit organisations sued five American AI developers on 12 November, asking a court to stop them developing models further. The defendants are Anthropic, Google, Meta, OpenAI and SpaceX’s AI business.
The plaintiffs are the Alliance for Secure AI, the Irreplaceable Education Fund, Public Citizen and The Heat Initiative. Their central claim is not that the technology failed. It is that the companies knew the risk and accelerated anyway, and that they have started handing research and development work to the systems themselves.
The argument the complaint makes
Brendan Steinhauser, chief executive of the Alliance for Secure AI, put the case in a statement: executives have spent years telling the public that their own frontier models could carry catastrophic risk, that the systems do not always behave as their developers intend, and that no reliable method exists to control them. The companies kept building and selling more capable systems regardless.
The filing goes after a specific mechanism. As developers come to rely on existing models to build the next generation, a process the complaint calls recursive self-improvement, the risk compounds because the oversight loop narrows with each iteration.
The plaintiffs also quote the defendants back at themselves. In 2015 Sam Altman said AI would probably, in the end, lead to the end of the world. In 2025 Dario Amodei put the probability of a catastrophic outcome at 25 per cent. Neither statement was made under oath. Both are on the record.
The conduct cited alongside the words
Public concern sharpened this year after a researcher left Anthropic and wrote publicly that the company and OpenAI were risking lives. The post drew more than 170 million views.
In September Amodei proposed that the industry coordinate to slow the rate of capability improvement at the frontier, buying time to verify safety. Altman and Musk both voiced support. The proposal was voluntary and unenforceable, and it did not change the shipping schedule.
Anthropic is proceeding with a listing on Nasdaq on 19 November at a target valuation of up to USD 1.4 trillion. On 9 November it released Claude Fable 5.2 and Claude Mythos 5.2, its most advanced models, alongside enterprise services aimed at financial and other professional users. It also signed a six-year compute purchase agreement worth USD 13.7 billion with the infrastructure firm RUM Group.
Musk endorsed slowing frontier development and did not leave the race. On 10 November he said pretraining had finished for Grok 4.8, a 2.5 trillion parameter model now in reinforcement learning.
OpenAI reached a multi-year strategic agreement with Firmus, an Australian AI infrastructure company, to deploy dedicated compute at an AI factory in Malaysia.
A widening set of actions
This is not the first legal challenge of the year. In June Florida attorney general James Uthmeier sued OpenAI over alleged ChatGPT safety failures, including information given to people connected to a school shooting, self-harm guidance, and dependency among younger users. In September Florida asked the court to bar OpenAI from developing new models without outside oversight.
In October eight more US states filed separately, alleging ChatGPT steered users, minors among them, towards harming themselves or others, and seeking penalties plus restrictions on developing self-improving models unless the company can show the systems align with human values.
A further nonprofit action in September targeted alleged intrusion by an OpenAI model into third-party computer systems on Hugging Face.
What a court can and cannot settle
The relief sought is an injunction against development, which is a hard remedy for a judge to grant against products that underpin public services, national security work and a listed company’s valuation. Courts are more comfortable with disclosure, audits and damages than with stopping research.
The more durable effect may be evidentiary. Every public warning an executive gave is now part of a pleading, filed next to a launch date, a compute contract and a valuation. That juxtaposition is difficult to argue with on the merits, because the two halves come from the same companies.
Regulation has moved slowly because legislators lack the technical grounding to write it. Litigation needs only a plaintiff with standing and a record to quote. The record now exists, and it was written by the defendants.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1101095940_115479.