AI Giants Were Sued for Agreeing to Slow Down—Subscribers Say Safety Became Collusion

Four of the world’s most powerful artificial-intelligence companies are being sued over an idea that sounds, at first, like the cautious option: slow down.

A proposed class-action lawsuit filed Friday, September 18, in the U.S. District Court for the Northern District of California accuses Anthropic, OpenAI, SpaceXAI and Google of reaching an unlawful agreement to reduce the pace of AI development. The plaintiffs are four people who pay for ChatGPT, Claude, Grok or Gemini subscriptions, and they argue that coordination among the rival labs would leave customers with less innovation and less value.

The case turns a familiar technology debate inside out. AI companies have faced years of pressure to prevent models from escaping human control, enabling cyberattacks or otherwise causing severe harm. Now, according to the complaint described by The Associated Press, some customers say a private safety pact among competitors could itself break the law.

The accusation is unproven. Filing a complaint is only the opening step in a civil case, and no court has found that the companies made an illegal agreement. Representatives for the four defendants had not responded publicly to AP’s request for comment as of Saturday.

The lawsuit focuses on one day

The plaintiffs point mainly to a burst of public statements on September 12. Anthropic CEO Dario Amodei had published a plan urging leading AI labs to coordinate on slowing certain advances while strengthening safety work. His proposals included giving independent evaluators unusually deep, continuing access inside frontier laboratories and involving democratic governments in cross-company discussions.

Amodei also recognized the antitrust risk. He suggested that the federal government might need to mediate the talks or grant a narrow legal waiver for specific safety conversations.

Other prominent executives then expressed support for the direction. OpenAI CEO Sam Altman backed a federal safety framework and said companies did not need to wait for new legislation or an antitrust exemption before beginning confidence-building work. Elon Musk, whose SpaceXAI operates Grok, wrote “Dario is right.” Google DeepMind co-founder and chair Demis Hassabis said the details still needed work but supported the direction.

An AP review of the executives’ positions also shows that the industry is not united. Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang have argued that companies should manage safety independently rather than through a coordinated slowdown.

Agreement is the legal fault line

Competitors are generally free to reach the same conclusion on their own. Several companies can independently decide that a product is too risky, that testing should take longer or that a model should not be released. Antitrust trouble begins when rivals make an agreement that restrains competition.

That means public expressions of support will not automatically prove the plaintiffs’ case. They will need to show that the defendants formed an actionable agreement, not merely that executives praised overlapping safety principles. They also must establish that paying subscribers suffered the kind of injury antitrust law recognizes.

The companies, if the case proceeds, can be expected to distinguish shared safety standards from coordinated limits on output. Industries routinely cooperate on cybersecurity, technical standards and emergency response. The harder question is whether a plan to make competing products advance more slowly functions like a safety standard—or like rivals agreeing not to compete as aggressively.

Why the subscribers say they are harmed

The lawsuit’s consumer theory is straightforward: subscribers pay recurring fees partly because the services are expected to improve. If the leading providers jointly restrain development, the plaintiffs say, customers would receive weaker products than open competition would otherwise produce.

But the safety argument cuts the other way. A faster release schedule may deliver new capabilities sooner while also increasing the chance that inadequate testing misses dangerous behavior. The companies can argue that monitoring, external evaluation and common safeguards improve the quality of the service rather than reduce it.

The dispute therefore is not simply “fast AI versus slow AI.” It is about who gets to set the pace, what competitors may discuss with one another, and whether government supervision is necessary before the labs coordinate.

Washington is moving in the opposite direction

The lawsuit arrives as the Trump administration resists broad new AI regulation. President Donald Trump said Saturday that he intends to appoint another AI adviser and create an “AI force,” but offered few details, according to Reuters. His administration has emphasized winning the technology race with China and has rejected calls for a sweeping federal slowdown.

That political posture makes a government-supervised industry pact less certain. It also raises the stakes of the California case: without a clear federal framework, a judge may be asked to draw an early boundary between permissible safety cooperation and anticompetitive coordination.

What happens next

The defendants will have a chance to answer the complaint and could ask the judge to dismiss it before evidence-gathering begins. The plaintiffs also must persuade the court to certify a nationwide class of paid subscribers.

Even if the case never reaches trial, it has already exposed an uncomfortable tension. Society wants AI companies to share enough information to prevent catastrophic failures. Consumers and regulators also want those same companies to compete. The lawsuit’s central question is whether the labs can build a common brake without turning it into a cartel.

Editor’s note: The complaint contains allegations that have not been proven in court. Anthropic, OpenAI, SpaceXAI and Google had not publicly responded to AP’s request for comment when its report was published.