China's Supreme People's Court released a national judicial policy document on 7 September, setting consistent rules for how courts should determine liability when artificial intelligence causes harm. The Opinions on Adjudicating Disputes Involving Artificial Intelligence in Accordance with the Law establishes fault-based liability as the default standard, while leaving key copyright questions around AI-generated content and training data unresolved for now.
The 24 provisions cover personality rights, personal information, intellectual property, consumer protection, autonomous vehicles and the use of AI in judicial proceedings. The opinions do not create new law. They guide courts in applying existing statutes - including the Civil Code, Personal Information Protection Law and Copyright Law - to AI disputes. China has not yet enacted a comprehensive AI law, and the SPC chose a flexible policy document over formal legislation or judicial interpretation, allowing the framework to evolve as technology and legal questions develop.
Chinese courts concluded 908 cases involving data ownership and transactions in 2025 alone, a 25.6% increase from the previous year. The SPC disclosed in April that it was accelerating drafting as courts confronted disputes involving AI-generated content, model parameters and data.
Fault liability becomes the default rule
The opinions make ordinary fault liability under Article 1165(1) of China's Civil Code the general rule for AI-related torts unless existing law provides otherwise. The SPC said this approach aims to avoid imposing excessive liability while AI technology and the industry remain at an early stage of development, which could discourage innovation.
Under this standard, the plaintiff generally must establish the defendant's fault. When determining fault, courts should consider AI-specific factors: the system's autonomy, technological transparency, potential risks, measures taken to prevent harm and the parties' ability to foresee and control those risks.
The U.S. has no comparable nationwide tort standard for the AI industry. Courts apply existing state doctrines case-by-case. In March, a California jury found Meta negligent in designing Instagram and liable for harm to a young user. China's approach differs by making ordinary fault liability the nationwide default for AI-related torts.
Responsibility rests with people, not AI systems
The opinions do not treat AI itself as a responsible legal actor. Responsibility rests with the people or entities involved in developing, providing or using AI systems, depending on their conduct and applicable law. The fact that harmful content was generated through AI does not itself relieve them of responsibility.
If a generative AI provider receives proper notice that its service is generating content infringing another person's personality rights and fails to take necessary measures promptly, the provider may be liable for resulting harm. Users who deliberately use infringing prompts to induce harmful output may also be liable.
For AI model training, processing personal information that individuals have made public, or that has otherwise been lawfully made public, generally will not constitute infringement if the processing remains within a reasonable scope and the individual has not expressly objected. Consent must be obtained where processing significantly affects an individual's rights and interests. Courts determining reasonable scope should consider the purpose of processing, its relationship to the model's functionality, the sensitivity of the information, potential effects on the individual and the uses the individual could reasonably have anticipated.
Copyright questions left for another day
In copyright infringement cases, a developer asserting noninfringement may be required to produce evidence concerning training-data sources, training records and model operations. This gives courts a mechanism to address evidentiary problems created by opaque systems where the provider controls the evidence needed to understand an AI system.
For companies, documentation of training-data provenance and model development may become important not only for compliance but also for litigation. However, the SPC left two consequential copyright questions unresolved: whether AI-generated content can itself qualify for copyright protection and how unauthorized use of copyrighted works for model training should be treated. The SPC acknowledged significant disagreement over these issues during drafting and concluded that further judicial experience is needed.
AI in the courtroom
The opinions also address generative AI use by lawyers and litigants. Parties using AI to prepare court filings or case-research reports must verify the authenticity and accuracy of cited laws, judicial interpretations and cases, disclose their use of AI and remain responsible for submitted materials.
Using AI to fabricate evidence or create false litigation may result in fines, detention or, where the conduct constitutes a crime, criminal liability. Courts are also given guidance for evaluating AI-generated evidence. The principle is straightforward: AI may assist people in judicial proceedings, but it does not replace human responsibility.
The SPC calls for coordination among courts, cyberspace regulators, public security authorities, procuratorates and market regulators, seeking a broader system combining industry self-regulation, administrative supervision and judicial protection.
Why this matters for legal professionals
The opinions signal that how a company manages training data, implements safeguards, maintains technical records and responds to infringement complaints may later affect a court's determination of liability - well before a dispute reaches court. For legal teams operating across jurisdictions, China is now adding nationwide judicial guidance to a global patchwork where the EU relies on comprehensive legislation and the U.S. depends on existing laws and case-by-case litigation. These differences matter for assessing cross-border AI liability, and the SPC's choice of ordinary fault liability as the default standard represents a deliberate attempt to balance innovation with accountability while the technology matures.
Your membership also unlocks:
