NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China's Top Court Wrote 24 Rules for AI. It Skipped the Two Everyone Argues About.

The Supreme People's Court's new AI opinion tells model developers to hand over training records, gives open-source projects a shield, and deliberately says nothing about whether AI output can be copyrighted.

2026-11-19 · 1171 words · NeuroAI
China's Top Court Wrote 24 Rules for AI. It Skipped the Two Everyone Argues About.

Imagine a company uses a generative model to make a short drama. Another company copies it, renames it, and runs ads against it. The first company sues.

The judge now has a document to reach for. And the most interesting thing about that document is what it declines to answer.

On September 7, China's Supreme People's Court published its Opinion on the Trial of Disputes Involving Artificial Intelligence (《关于依法审理涉人工智能纠纷案件的意见》), issued as Fa Fa [2026] No. 10. It runs to five parts and 24 articles, and the court describes it as the first judicial rules document on AI issued by a country's highest court.

Two subjects that dominate every AI copyright argument — whether machine output can be copyrighted, and whether training on other people's work is fair use — are not in it. The court says so explicitly.

Key takeaways

  • The opinion sets out how courts should allocate liability among AI developers, service providers and users when generated content infringes copyright, and requires developers claiming non-infringement to produce their training-data sources and training records.
  • It gives open-source developers a path to liability exemption where they disclose a module's function and security risks.
  • It is not a new AI law: it is a set of adjudication guidelines telling lower courts how to apply the Civil Code, the Copyright Law and other existing statutes to AI cases.
  • It deliberately leaves two questions open: whether AI-produced content is copyrightable, and how to characterise training on copyrighted works. The court says consensus is not yet there.

What the 24 articles actually do

The opinion is not a new AI law. China still does not have one. It is a set of adjudication guidelines telling lower courts how to apply the Civil Code, the Copyright Law, the Personal Information Protection Law and other existing statutes to AI cases.

Its five parts cover principles, torts, intellectual property, evidence and court machinery. Part two is the widest: deepfaked faces and voices, "resurrecting" deceased people as digital avatars, doxxing carried out with AI, training on personal information, provider liability after notice, product liability, algorithmic price discrimination and assisted-driving crashes. Part three is the block most relevant to anyone building with generative models.

The burden of proof is the real change

Article 12 is where the practical weight sits.

When AI-produced content is alleged to infringe copyright, the court must weigh the type of AI service, where the training data came from, how much each party contributed, what preventive measures were taken, and who profited. That is a balancing test, not a strict rule.

But the procedural move is sharper. If an AI developer argues non-infringement, the court must order it to produce the source of training data, training records, the model's operating mode, and the scientific basis for its claim. The court's explainer notes that this evidence is not something a rights holder can obtain alone.

For a rights holder, that is a meaningful shift. You no longer have to prove what is inside the model. You have to raise a plausible case that the output is substantially similar, and the developer then has to explain itself. For a model company, training-data documentation is no longer a governance nicety. It is litigation infrastructure.

The open-source shield

Article 13 addresses a question that has nagged at Chinese open-weight releases: if I publish a model or a code module and someone uses it to infringe, am I liable?

The answer is a qualified no. Courts should weigh the type of open-source licence, its restrictions, the security measures taken and how much was disclosed. Where a developer releases a module for free, publicly explains its function and risks, and a third party's use causes infringement, the court may find the developer not liable.

That is not blanket immunity — the discretion leaves room — but it is a clearer safe harbour than open-source projects had before.

The two blanks

The court says that during drafting, views diverged sharply on two issues: whether AI-produced content can be copyrighted, and how to characterise unauthorised use of works to train large models (大模型). It therefore makes no rule on either, and says it will revisit when experience accumulates.

In practice, an AI copyright case in China will still turn on the facts of human involvement — how prompts were written, how outputs were selected, how much editing happened — rather than on the medium.

The case that shows how it lands

A court in Wuhan's Jiang'an District illustrated the point in September.

A company produced a 47-episode microdrama (微短剧) using generative AI tools, filed with the National Radio and Television Administration and published on short-drama platforms. A second company copied the series, changed the title, and monetised it with ads.

The court held that the drama was a protected audiovisual work. The reasoning was not that AI made it, but that people did: script planning, shot-by-shot prompt design, asset selection, filtering of generated clips, editing, sound and subtitle work. The AI was a tool inside a human process, and the humans controlled the result.

Damages were RMB 20,000 (≈ US$2,800). The court factored in the compute cost of tokens consumed and the licensing cost of commercial tools — an attempt to price AI-specific inputs into the remedy.

That is a small award, and a first-instance decision that could be appealed. But it shows the direction: protect the human layer, and treat the machine as equipment.

Honest limitations

I have read the opinion and the court's explanatory Q&A. I am not a lawyer; this is a description of the document, not legal advice.

The opinion binds Chinese courts, not regulators, and its effect depends on how lower courts apply it. The discretion written into the open-source article is exactly the kind that produces divergent outcomes.

The Wuhan case is a first-instance district court judgment reported by state media. I have not seen the full judgment or any appeal, so the damages figure is one data point, not a tariff. And the court's silence on the two open questions is deliberate but not permanent: anyone building a business on AI-produced content in China should assume the rule will be filled in.

What readers can do now

  • If you train or fine-tune models for the China market, document your data now. The opinion turns training records from a compliance checkbox into something a court can order you to produce.
  • If you publish open weights, publish the risk notes too. The Article 13 exemption leans on disclosure of function and security risk. A README that says nothing is a weaker position than one that does.
  • If you make creative work with AI, keep the human trail. Prompt drafts, selection notes, edit history and version files are what turn "the model made it" into "we made it with a model".

The court has drawn a line around the argument without settling the argument itself. For now, the safest asset in an AI copyright dispute is not the model. It is the paper trail showing a person made choices.

Related coverage

Cite this article

NeuroAI (脑机网), "China's Top Court Wrote 24 Rules for AI. It Skipped the Two Everyone Argues About," 19 November 2026, https://neuroai.site/a/na-aigc-supreme-court-ai-rules

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Generative Media” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories