Supreme Court Draws a Red Line for AI: How the 24-Article Opinion Defines the Boundaries of AI-Generated Content

On September 7, the Supreme People's Court issued the Opinions on the Trial of Cases Involving Artificial Intelligence in Accordance with the Law (Fa Fa [2026] No. 10). Divided into five parts and 24 articles, the document is China's first systematic judicial adjudication document specifically addressing AI disputes. Over the past few years, AI-related cases have piled up in courts across the country: AI face-swapping, commercial use of AI voices, data scraping for model training, defamation through large-model hallucinations, copyright ownership of AI-generated content, and accidents involving intelligent assisted driving. Each category has produced precedents, but the standards are inconsistent, and judges have mostly been applying traditional civil law to new scenarios.

What this opinion does is set a unified coordinate system for these scattered precedents.

AI Is Not a Legal Subject: The Starting Point of the Entire Logic

Article 2 states it plainly: AI itself does not have the capacity to be a civil subject, and the relevant legal liability is borne by civil subjects such as developers, service providers, and users.

This position is not new. In the country's first AI hallucination infringement case heard by the Hangzhou Internet Court, the court held that a "promise of compensation" automatically generated by AI did not have legal binding force because there was no responsible subject with civil capacity behind it. The opinion elevates this individual case into adjudicative logic applicable nationwide.

Once the question of subjecthood is settled, the rest becomes easier: models are tools; when something goes wrong, you look for the person, not the model.

Three Pillars: Human-Centered, Innovation-Supportive, and Safety Guardrails

The opinion establishes three principles: adhering to a human-centered approach, supporting innovative development, and strengthening safety guardrails. Read together, they reveal the regulatory orientation: there is no intention to trap AI with a single yardstick, but technology will not be allowed to run ahead of the rules.

When translated into specific provisions, the human-centered principle focuses on personality rights and personal information. High-frequency disputes such as AI face-swapping, AI voice cloning, and impersonating celebrities to sell products are all explicitly named. The opinion stresses the need to make full and good use of existing laws, prevent algorithmic discrimination, and protect individuals' right to know and right to choose in automated decision-making.

The innovation-support principle is mainly reflected in defining the "reasonable scope" of training data. Previously, courts had inconsistent understandings of using public personal information to train large models, some more permissive and some stricter. The opinion does not take a one-size-fits-all approach, but instead provides more detailed directions for judgment.

Safety guardrails, by contrast, point more toward platform obligations and the allocation of infringement liability.

The Provisions Creators Should Care About Most

For those working in content, two areas in this opinion are the most practical.

One is the copyright of AI-generated content. The opinion continues the judgment path of "what the human did": if there is substantial human input and creative choice behind AI-generated content, there is room to discuss its status as a work. This line of thinking is consistent with two local judgments in September. In Shanghai's first AI voice infringement case, damages of 50,000 yuan were awarded, and in Wuhan's AI short drama copyright case, damages of 20,000 yuan were awarded; in calculating compensation, both courts took production costs such as token consumption into account. Whether tool costs can count as part of creative input had no clear answer in the past.

The other is the boundary of training data. The opinion requires clarifying the responsibilities of large models at each stage—data collection, fine-tuning, and content distribution—involving multiple dimensions such as reproduction rights, personal information rights and interests, voice personality interests, reputation rights, and the right to disseminate information online. In other words, throughout a model's life—collecting data, tuning parameters, answering questions, and distributing content—every stage may carry legal risk, not just "the moment of generation."

Relationship with Existing Laws

The opinion repeatedly emphasizes that it does not create new civil rights, but rather clarifies how existing frameworks—such as the Civil Code, Copyright Law, Personal Information Protection Law, Data Security Law, Anti-Unfair Competition Law, and Consumer Rights Protection Law—apply in AI scenarios.

The division of labor with the Interim Measures for the Administration of Generative Artificial Intelligence Services is also clear: the Measures govern the administrative regulatory level and impose obligations on service providers; the opinion governs the determination of civil infringement and serves as the basis for courts to decide cases. Administrative regulations cannot directly serve as the basis for civil adjudication; this gap had long existed, and the opinion fills it.

The opinion also identifies problem areas including big data price discrimination against existing customers, liability for autonomous driving accidents, and reputational infringement caused by AI hallucinations, and provides guidance on evidence rules. How to present evidence and trace AI-generated content has long been a practical difficulty.

What It Means for Businesses and Creators

From a compliance perspective, this opinion turns "after-the-fact remediation" into a "pre-event checklist." Teams building content products should at least review several things: where training data comes from and whether it is authorized; whether AI-generated content involves the likenesses and voices of real people; whether user-facing features provide notice of automated decision-making; and whether the content distribution stage has a channel for handling infringement complaints.

Overseas, answers were also emerging at the same time. In September, the U.S. Court of Appeals for the Third Circuit ruled in Thomson Reuters v. Ross Intelligence that using copyrighted material to train a competing commercial AI tool does not constitute fair use—the first time a U.S. federal appellate court has weighed in on the relationship between AI training and fair use. China uses a unified judicial opinion, while the United States uses case-by-case precedents. The paths differ, but the questions they point to overlap heavily: where is the boundary of training data, and who owns the rights to generated content?

Writing the rules is only the first step. Many of the 24 articles are statements of principle, and how they apply to each specific case will still need to be filled in by subsequent judgments. For practitioners, what they can do now is self-audit the parts they can self-audit first, rather than waiting for a judgment to arrive at their door before catching up.