China's First AI Rulebook Is 24 Articles and Two Deliberate Blanks

TL;DR

On September 7 the Supreme People's Court of China published the country's first set of adjudication rules written specifically for AI disputes. Five parts, 24 articles. It covers face swaps, voice clones, doxxing, algorithmic price discrimination, self-driving crashes, and what happens when you hand a judge a brief your model invented.

The organizing idea is simple enough to fit on a sticky note: if the output is a recognizable replica of a real person and that person did not consent, you are liable. The interesting part is what the court left out. Two questions get no answer, and they are the two that every AI copyright case outside China is currently fighting over.

What this document is, and what it is not

The full title is the Opinion on Lawfully Adjudicating Cases Involving Artificial Intelligence Disputes. It was presented by SPC Vice President Tao Kaiyuan and framed around three stated principles: people first, support innovation and development, and hold a security line.

Get the category right before you read anything into it. This is not legislation, and it is not a judicial interpretation carrying statutory force. It is an Opinion: the top court telling every court beneath it how these cases should come out. Think of it less as a new law and more as head office mailing every branch the same decision tree. Nothing in the statute book changed, but every clerk now reaches for the same answer.

That still matters a great deal, because until this week Chinese lower courts were improvising from the Civil Code, the Personal Information Protection Law, and the 2023 Interim Measures for generative AI services, which say training must not infringe intellectual property and then stop talking.

Consent is the line, and it covers your voice

Part 2 runs from article 3 to article 11 and does the heavy lifting. The text reported by Chinese outlets holds that processing a person's name, likeness and similar identifiers with AI to generate a recognizable virtual digital image, then using or publishing it, infringes personality rights. The same logic is applied to voice: using someone's speech as training material to synthesize their vocal characteristics without consent infringes their voice rights.

Two extensions are worth noting because they close the obvious workarounds. Operating a synthetic avatar or cloned voice to make false statements that lower someone's social standing is treated as reputational harm, so you do not escape by arguing the clone was flattering. And using AI to imitate a celebrity endorsement in commerce is treated as fraud, which under Chinese consumer protection law opens the door to punitive damages rather than ordinary compensation.

Alongside those sit two provisions aimed squarely at consumer-facing systems. Algorithmic price discrimination, the practice Chinese users call big data killing the familiar, is actionable when a merchant applies unreasonable differential treatment on price or terms for the same goods and a consumer is harmed. And AI-assisted doxxing, locally known as opening the box, is regulated as a privacy violation.

Who is actually on the hook

Liability is not parked entirely on the person who typed the prompt. The Opinion assigns responsibility across the chain: generative AI service providers can be held liable where they are notified that their system produced rights-infringing content and fail to act in time, and both network service providers and AI service providers can be ordered to stop the infringement or prevent further harm. Users who deliberately steer a model toward infringing output carry their own exposure.

The court also gave itself a faster lever: where content threatens serious and irreversible harm, judges can issue an injunction rather than wait for the damage to become permanent and then price it.

The two deliberate blanks

Part 3, articles 12 through 16, is the intellectual property section: infringement determination, open source software obligations, patent authorization and confirmation, technical contract performance, and data use norms. Open source obligations getting an explicit slot is notable on its own, given how many model releases ship under licenses nobody has litigated.

What is not there is the headline. The Opinion does not resolve whether AI-generated content attracts copyright, and it does not resolve liability for training a large model on unlicensed works. Both were reported as consciously held back, on the reasoning that consensus has not formed and the questions need more judicial exploration first. Tao Kaiyuan's framing was that the Opinion cannot cover every case type and that many issues require further accumulation.

Read that next to the docket elsewhere. Those two questions are the entire substance of the author and publisher suits currently grinding through US courts. China's top court looked at the same two questions, wrote 24 articles around them, and left them blank on purpose. Whatever else that is, it is not an accident.

The provision that will bite lawyers first

Part 4, articles 17 through 20, handles fact-finding and evidence review, and it puts an explicit verification duty on parties who submit AI-generated litigation documents and case-law search reports. Courts are also directed to punish the use of AI to fabricate materials or obstruct proceedings.

This is a polite way of saying the bench has read enough invented case citations for one lifetime. It is also the single most portable idea in the document: the obligation attaches to the human who filed, not the tool that drafted. If you are building anything that generates a document a professional then signs, that is the liability shape you are designing around.

What a builder should actually change

• Treat voice as biometric-grade. If your pipeline ingests real speech to produce a synthetic voice with that speaker's characteristics, consent is now the explicit test in a market of a billion-plus users. Store the consent artifact, not just the audio.
• Build the notice path before you need it. Provider liability turns on failing to act after notification. A reachable takedown channel with a logged response time is cheap; discovering you did not have one is not.
• Audit your pricing logic. Differential pricing on identical goods without a defensible basis is named as actionable. If you cannot explain your price surface in one sentence to a judge, that is a finding.
• Check your open source obligations. The IP section calls them out by name, which means somebody is going to be the test case.

Caveats, stated plainly

There is no official English translation of the Opinion yet, and the article-level readings above come from Chinese-language reporting on the released text, principally Jiemian, China News Service and IT Home, cross-checked against Xinhua's English summary. One point rests on a single outlet: Jiemian reports that training on already-public personal information generally does not require fresh consent unless the information significantly affects personal interests, which tracks the existing Personal Information Protection Law carve-out. Treat that one as reported rather than settled.

An Opinion is also guidance, not a statute. Lower courts follow it, but how it behaves once real defendants start litigating around the edges is the part nobody knows yet.

Key Takeaways

• China's Supreme People's Court released its first AI adjudication rules on September 7: five parts, 24 articles, presented by Vice President Tao Kaiyuan.
• Consent is the dividing line for recognizable digital replicas of a real person, and that explicitly includes cloned voices, not just faces.
• Liability spans the chain: users who steer models toward infringement, and service providers who ignore a notice, both carry exposure.
• Nine of the 24 articles sit in the tort section, which also captures algorithmic price discrimination, AI doxxing and self-driving crash liability.
• Fake celebrity endorsements are treated as fraud, which unlocks punitive damages rather than ordinary compensation.
• Copyright in AI output and liability for training on unlicensed works were deliberately left unanswered, on the stated ground that consensus has not formed.

Sources: Xinhua, China News Service, China News Service (case types), Jiemian, Sina Finance, IT Home, TechXplore, Supreme People's Court of China