First: this is an allegation, not a judgment.
A proposed class action filed on September 18, 2026, in the US District Court for the Northern District of California alleges that Anthropic, OpenAI, SpaceXAI and Google illegally coordinated to slow AI development. The plaintiffs are four paying subscribers. At the time of the Associated Press report, the companies had not responded, and no court had found the alleged conduct or any violation.
The complaint is not simply arguing that AI must never slow down. The plaintiffs say they do not object to a company restraining itself for safety. Their question is whether competitors collectively substituted a private restraint for individual accountability and market choice.
The moment a company invokes safety, its brand makes a promise.
Safety is an attractive word. It is hard to oppose. That is why it deserves more scrutiny, not less. When a company says, “We are slowing down to protect everyone,” it moves beyond a technical claim. It makes a brand promise.
That promise can be costly. A company may delay a product it could ship, leave revenue on the table and ask competitors to exercise the same restraint. But sacrifice alone does not prove public benefit. The company still has to show who defined the risk, what evidence supports the threshold and whether the rule applies when it is inconvenient to the company itself.
A brand is not built by saying it is safe. It is built by showing why it stops when it could move faster.
When “slow down together” becomes an entry barrier.
Dario Amodei has proposed a three-part approach for pacing frontier AI: third-party evaluation, democratic coordination and international coordination. He also acknowledges that voluntary coordination among companies raises antitrust problems and suggests government mediation or a narrow legal waiver.
That is the tension. Shared standards may reduce risk. If those standards require expensive evaluation and scarce compute, they may also keep smaller rivals out. When a gate built in the name of public benefit is sized so only incumbents can pass through, safety stops being a brand virtue and becomes market power.
Uncertainty is not a license to race—or to collude.
The International AI Safety Report 2026 documents harms from false information and reasoning failures and identifies reliability as especially important for agents. It also says current agents do not yet have the sustained autonomous capabilities assumed in loss-of-control scenarios, while real-world evaluation remains incomplete.
Uncertainty does not justify moving at any speed. It does not justify letting a small group of companies privately set the speed of a market either. From a brand perspective, what is missing is not a louder pledge but clearer evidence: the threshold for stopping, who it covers, the cost, the conditions for restarting and a way to challenge the decision.
What a company invoking safety should be prepared to prove.
- Does it apply the same threshold when the rule is costly to itself?
- Can outsiders understand why development stops and what permits it to restart?
- Do competitors and new entrants receive comparable rules and access to evaluation?
- When the judgment is wrong, who corrects it and explains what changed?
Choosing either safety or competition is the easy question. The harder work is protecting safety without allowing power to pool in private hands. Part II examines how independent evaluation, public thresholds and a route for challenge can turn a pledge into a structure that can be verified.
Scope and limitations.
This is a brand and governance analysis based on reporting about the public complaint and other public materials. It does not determine that any defendant violated the law and is not legal advice. Allegations and company proposals may change as responses, evidence and judicial decisions emerge.