China Wrote the First Rules for AI Lies. Proving One Happened Is Still on You.

2026-09-13

On September 7, China's Supreme People's Court published the first national judicial rules anywhere in the world dealing squarely with AI-generated harm: deepfakes, cloned voices, algorithmic price discrimination, and chatbot hallucinations. Five parts, twenty-four articles, and a mechanism that sounds almost administrative next to the scale of what it governs. A service provider isn't automatically liable when its AI system invents something false about a person. It becomes liable once it has been notified and fails to act.

Notice and act. Two words that quietly move the entire burden of the problem onto whoever was harmed.

Courts elsewhere are still arguing about something more basic: whether an AI system's output counts as speech at all, and if so, whose. In Delaware, a defamation claim against Google over what its AI system said about a public figure only reached discovery this summer, after months of argument over that exact question. The European Union's AI Act, whose transparency obligations became enforceable across the bloc in August, requires systems to disclose that a person is talking to AI. It says nothing about what happens after a system has said something false and someone objects.

China's rule set skips the philosophical fight and goes straight to procedure. Notify the provider. Give them a chance to fix it. If they don't, they're on the hook. It is, on paper, a cleaner and faster path to accountability than anything currently on the books in Washington or Brussels.

The rule assumes a kind of proof that doesn't naturally exist

Here is the gap nobody has written a rule for yet: what does a valid notice actually require you to hand over?

A chatbot's answer is not a fixed document. Ask the same question twice, minutes apart, and a system can return two different responses, shaped by a prompt variation, a model update, or nothing you can identify at all. There is no original to subpoena, no server log a claimant can access, no guarantee the exact wording that harmed someone will ever reappear once they go looking for it. A screenshot proves someone captured something on some screen; it does not prove a date, a source, or that the image wasn't altered after the fact. None of that is a problem the Supreme People's Court's twenty-four articles were written to solve. They describe what a provider must do once notified. They say nothing about what a claimant must produce to notify credibly in the first place.

That silence is not a drafting oversight so much as a preview of where every jurisdiction working on this is headed. The EU is reportedly weighing a liability annex to sit alongside its transparency rules. Several US state legislatures have hallucination-specific consumer protection bills in committee. Each is approaching the same problem from a different legal tradition, and each is going to run into the identical procedural hole: a system with no obligation to remember what it said, answering a question that may never be asked in exactly the same way twice, about a person who has to prove, after the fact, what happened at a moment already gone.

Different courts, same missing piece

What makes this moment interesting isn't that China moved first. It's that the direction of travel is now visible from three different starting points at once, and all three arrive at a version of the same requirement: someone has to be able to show, credibly and after the fact, what a given AI system said, when it said it, and where that answer plausibly came from. Not a recollection. Not a screenshot taken under uncertain conditions. Something closer to a record — timestamped, attributable, and built before the dispute started rather than assembled to win it.

The legal theory of AI liability is arriving faster than the evidentiary infrastructure to support it. China's court just made that gap explicit, in writing, for the first time. Whoever is first affected by a hallucinated claim under this new standard — a company, a public figure, a legal team weighing whether to send that notice — will discover the actual bottleneck isn't the law. It's whether they can prove, on the day it mattered, what the machine actually said.