China's Supreme People's Court issued nationwide judicial guidelines on artificial intelligence disputes on September 7, 2026, drawing the world's first detailed legal line on AI harm. For legal professionals, the rules answer a question every jurisdiction has been dodging: when a chatbot lies, a deepfake scams someone, or an algorithm discriminates, who exactly pays and under what standard.
The guidelines, formally titled the Opinion on Properly Hearing Cases Involving Artificial Intelligence Disputes, establish that three parties can bear responsibility for AI-generated harm: the person who created the fake, the platform that hosted it, and the user who typed the prompt. No other major economy has gone this far in spelling out liability.
Deepfakes as theft
The court treats a cloned face or voice as theft the moment someone builds it without consent. You cannot construct a talking avatar from someone else's likeness simply because the tools make it easy. If you use that clone to spread a false claim or damage a reputation, you face liability twice - once for the unauthorized copy and once for the harm caused by distributing it.
This framing arrives amid a surge in deepfake-driven fraud. A finance worker in China was tricked into wiring 1.86 million yuan, roughly $262,000, after a video call with what looked and sounded exactly like her boss. It was a clone built from public video and a cloned voice track. Hong Kong police reported a separate incident in which a multinational lost approximately $26 million through a faked video conference call.
The hallucination problem and the notice-and-act standard
Generative models fabricate quotes, invent details, and describe events that never occurred. Under the new rules, a company running such a model is not automatically liable for a hallucination it did not know about. But once someone flags false, damaging output and the provider does nothing, the clock starts. Failure to act means the platform shares blame with whoever prompted the model.
Zhou Jiahai, head of the Supreme People's Court's research office, explained the reasoning: "We cannot expect every consumer to become an expert at spotting deception. The law must step in promptly to protect consumers' legitimate rights and interests."
This notice-and-act standard borrows the logic long used for copyright infringement online. Platforms are not liable for what users create. They are liable for what they leave up after being told about it. For companies like Baidu, Alibaba, and ByteDance, all of which run consumer-facing generative AI products inside China, this is a familiar compliance burden. Takedown desks that once handled pirated video and counterfeit listings must now handle deepfakes and hallucinated defamation too.
A user who deliberately prompts a model to produce something defamatory or fake, knowing it is false, cannot hide behind "the AI said it." The guidelines place intent squarely on the person at the keyboard.
Algorithmic price discrimination
The court also addressed a quieter harm: algorithmic price discrimination, sometimes called "big data killing familiarity" in China. This occurs when a platform charges loyal or frequent customers more than new ones for the identical product. The guidelines say merchants who use algorithms to set different prices or terms for the same good, without reasonable justification, can be held liable for the difference.
Global context
The European Union's AI Act sets risk categories and disclosure duties but does not spell out, case by case, who pays when a model hallucinates a defamatory claim. The United States has no federal deepfake liability statute, relying instead on a patchwork of state laws and ordinary defamation claims stretched to cover a new kind of harm. Regulators in Washington and Brussels have spent years debating hypotheticals that Beijing's top court just answered with actual liability rules.
Vice President Tao Kaiyuan described the guidelines as an effort to balance development and security, leaving room for interpretation as unanticipated AI harms emerge. That phrasing is deliberately loose. Twenty-four provisions cannot anticipate every way a model will misbehave next year. Judicial guidelines still depend on judges applying them consistently across thousands of local courts.
Why this matters for legal professionals
The framework establishes concrete liability chains that will influence how courts in other jurisdictions approach AI harm. For lawyers handling cross-border technology disputes, these guidelines provide the first detailed judicial reasoning on apportioning blame among creators, platforms, and users. Paralegals and legal researchers working on AI for Paralegals cases will encounter these standards as comparable models spread. The notice-and-act mechanism, in particular, creates a compliance template that platforms operating globally may adopt preemptively, regardless of whether their home jurisdictions have equivalent rules. The question is no longer whether AI harm creates liability. It is which party's balance sheet absorbs it.
Your membership also unlocks:
