Nobody Passed an AI Law Last Week. Four Different Bodies Wrote the Rules Anyway.

In the last two weeks the working rules for AI changed in the United States, China and at the G20. Not one of those changes was a law.

Here is what actually happened.

The US Justice Department filed a 20 page statement of interest in the New York Times case against OpenAI and Microsoft, telling a federal court that training large language models on copyrighted text is generally protected by fair use. Four days later the Seattle Times and Newsday sued the same two companies over the same practice anyway.

China's Supreme People's Court announced on September 6 that it would release, the next day in Beijing, its opinions on how courts should handle disputes involving artificial intelligence.

At the G20 innovation ministerial in Chapel Hill, ministers adopted the nonbinding Carolina Principles, which favour flexible sector by sector oversight over new horizontal AI laws. The substance of that document is a decision not to write rules.

And a single US district judge vacated the Pentagon's designation of Anthropic as a supply chain risk, calling it unconstitutional retaliation for the lab's refusal to drop safety limits on military use.

Four instruments. A brief, a set of judicial opinions, a communique, and one ruling. Zero statutes.

Which of the four actually binds anybody? Only the ruling. A statement of interest is an argument the court is free to reject. Judicial opinions from China's top court are guidance rather than legislation. The Carolina Principles are nonbinding by design.

That is the part worth sitting with, because the ranking by legal force is close to the reverse of the ranking by practical effect.

Guidance to judges reaches every lower court that will hear one of these cases, and it arrives before most of those cases are filed. An executive branch brief tells every future defendant which argument the government will back. A nonbinding communique with the US pushing it, and reportedly China signing it, sets the ceiling on what any national regulator will attempt for the next couple of years. The one binding item, the Anthropic ruling, applies to exactly one company.

The clock is the story. Legislatures work in years. Courts work in months. A procurement decision, or a blacklist, works in weeks.

Companies plan around the fastest clock, which means the venue is starting to matter as much as the model. That is my honest read of where this goes over the next year: jurisdiction becomes a product decision, the way data residency did a decade ago, and the question shifts from which model you deploy to where a dispute about it would be heard.

There is a second turn here that almost nobody is watching. The bodies writing these rules are also becoming users. Japan's Supreme Court has put roughly 60 million yen into its fiscal 2027 budget request to test AI tools in civil trials. India's chief justice describes AI already running inside the judiciary for legal research and judgment translation into 16 languages, while insisting it cannot replace judicial conscience. The venue is becoming a customer.

The caveats are real. A statement of interest can be ignored by the judge it was filed with. We have China's announcement and briefing, not yet a careful reading of what the opinions say. The EU AI Act is still a law and still applies. And two weeks of filings is not a system.

What to do with this. Two things.

Track the judicial calendar, not the legislative one. If you deploy across borders, the next fact that changes your exposure is far more likely to be a court's guidance or a procurement designation than a bill. Ask your counsel which of the four instrument types above actually reaches you, and you will probably find the answer is the two nobody sent you a briefing note about.

Then watch one falsifiable thing. If the next wave of US copyright filings comes from publishers with mastheads, this stays a commercial negotiation conducted through courts. If it comes from individual rightsholders without a masthead behind them, the economics change, because the settlements that clear a newspaper's claim do not scale to a million people. The Seattle Times filing four days after the government said the practice is lawful suggests nobody is waiting for the policy question to be settled first.