The Pulse
China’s Top Court Sets National Rules for AI Disputes
China’s Supreme People’s Court has issued a 24-article opinion directing courts nationwide on disputes involving deepfakes, voice cloning, AI hallucinations, consumer discrimination and autonomous driving. The guidance applies existing civi
AI.info Team ·
China’s Supreme People’s Court has issued the country’s first national judicial guidance dedicated to artificial-intelligence disputes, a 24-article document that tells courts how to handle deepfakes, cloned voices, AI-generated misinformation, algorithmic price discrimination, model training, copyright claims and autonomous-driving accidents.
Published on September 7, the opinion does not create a standalone AI statute. Instead, it directs courts to apply China’s existing Civil Code, Cybersecurity Law, Data Security Law, Copyright Law, Anti-Unfair Competition Law, Consumer Rights Protection Law, Personal Information Protection Law and Civil Procedure Law to disputes involving AI.
The document gives courts a common framework for assigning responsibility across developers, service providers and users. It also sets out rules for evidence, false litigation and criminal conduct involving AI. For companies building or deploying models, the guidance turns a series of scattered court disputes into a national set of judicial instructions.
Twenty-four Articles Cover Deepfakes, Data and Driver Assistance
The opinion is divided into five sections covering general principles, infringement disputes, intellectual-property cases, litigation procedures and court-management mechanisms. The court says its aim is to balance innovation with safety, while ensuring that people retain legal protection when AI systems affect their identity, privacy, reputation, property or physical safety.
Courts are instructed to begin with three principles: put people first, support innovation and maintain a safety boundary. The document says liability should reflect the type of AI system involved, the harm caused and each party’s ability to control the relevant risk. It distinguishes between general-purpose and specialized models, as well as open-source and closed-source systems.
China’s highest court also adopts fault-based liability as the default approach for AI-related civil infringement when existing law does not specifically require strict liability or a presumption of fault. That provision matters because it rejects a blanket rule that would automatically make every model developer responsible for every harmful output.
At the same time, the guidance does not treat AI as a legal black box. Courts may examine the conduct of developers, providers and users separately, including the precautions they took, the way a system was designed and the extent to which a party could prevent or reduce the harm.
AI Faces and Voices Need Consent
The opinion directly addresses the growing use of AI to reproduce a person’s appearance and voice. Using someone’s name or likeness without consent to create and publish a recognizable virtual digital image can infringe personality rights, according to the court. The same applies when a person’s voice is used as training material to imitate their timbre, intonation and speaking style in a recognizable synthetic voice.
Manipulated digital images and voices can also create liability for reputational harm. If an operator controls a virtual likeness or synthetic voice to spread false claims or lower another person’s social standing, the conduct may constitute an infringement of reputation rights.
The rules cover cases sometimes described as AI “resurrection,” in which technology is used to recreate the appearance or voice of a deceased person. The court’s framework focuses on recognizable identity and the rights affected by the use, rather than treating synthetic media as harmless because it is produced by software.
China has already seen courts deal with similar disputes. The Beijing Internet Court ruled in 2024 that using a voice actor’s recorded audiobooks to create a commercial text-to-speech product without consent infringed her voice rights. The new national guidance gives lower courts a broader set of principles for handling such cases.
Public Personal Data Can Be Used, But Not Without Limits
The court’s opinion attempts to draw a line between AI training and personal-information protection. It says that processing lawfully public personal information for model training within a reasonable scope will generally not be treated as an infringement when the individual has not expressly refused that use.
That permission is not unlimited. When processing has a significant impact on a person’s rights or interests, the opinion says the legally required consent must still be obtained. The approach gives developers room to use some publicly available information while preserving additional protection for high-impact uses.
The guidance also targets “doxxing,” including the use of AI to collect and analyze publicly available details such as phone numbers, online accounts and social-media profiles in order to obtain, expose or distribute private information. The court treats the use of public fragments to identify and reveal private data as a potential privacy infringement.
Consumer disputes receive similar attention. Businesses that use algorithms to impose unreasonable differences in prices or other transaction terms can face liability when consumers suffer harm. The rules also address AI-generated impersonation in marketing, including fake celebrity endorsements. Where such conduct amounts to fraud, courts may support claims for punitive damages.
Model Providers May Face Duties After Notice
The opinion allows courts to apply China’s existing notice-and-takedown framework to generative-AI services. If AI-generated content infringes another person’s reputation, privacy or other personality rights, a service provider that receives notice and fails to take necessary steps, such as stopping further generation of the material, may be held responsible for resulting harm.
The provision gives rights holders a route to challenge harmful outputs without first proving that the service provider created the original prompt or intended the result. It also places pressure on providers to maintain procedures for receiving complaints and responding to known violations.
Intellectual-property disputes receive a case-by-case approach. Courts are told to consider the type of AI service, the industry involved, the source of training data, the participation of each party, the safeguards adopted and the profits earned when deciding whether developers, providers or users should bear responsibility for copyright infringement.
Open-source software receives separate treatment. Courts should consider the open-source licence, restrictions on use, security and compliance measures and the extent of risk disclosures. A party that provides code modules for free, clearly explains their functions and discloses known security risks may be found not liable for infringement under the circumstances described by the opinion.
AI Evidence Gets Its Own Courtroom Rules
One of the document’s most practical sections concerns evidence. AI-related cases often involve technical facts that one side cannot independently verify, so the opinion directs courts to strengthen guidance for litigants and to expand judicial investigation when parties cannot obtain necessary evidence themselves.
Judges may rely on expert assistants, technical investigators, appraisers and jurors with relevant expertise. When a party controls important evidence and refuses without a valid reason to produce it, the court may accept the opposing party’s claim about what that evidence would show.
The opinion also sets specific review standards for common forms of digital evidence. Courts should examine the source of raw data, data-cleaning rules and analytical methods in big-data reports. For blockchain evidence, judges should check whether the data was authentic before being uploaded and whether the technical platform was reliable.
AI-generated material offered as proof requires additional scrutiny. Courts may consider the prompt design, the prompt’s effect on the output, the similarity between generated content and the work at issue, the consistency of repeated tests, the model’s training data, algorithm design and content-filtering systems.
The court warns that litigants cannot use AI to manufacture contracts, invoices, audio or video files, submit false claims or obstruct proceedings. People who use AI to prepare legal filings, case-retrieval reports or other court materials must verify that the laws, judicial interpretations and cases cited are real and accurate, disclose that AI assistance was used and accept legal responsibility for the submitted material.
China’s Courts Leave Some AI Questions Open
The guidance reaches criminal conduct as well. It instructs courts to punish AI-assisted fraud, insults, defamation, damage to commercial reputation, illegal acquisition of computer data, violations involving personal information and the production or distribution of illegal sexual material when those acts constitute crimes.
It also addresses drivers who activate assisted-driving systems and then evade safety monitoring with unauthorized devices or by abandoning their responsibility to control the vehicle. If an accident leads to a crime, the driver remains responsible. Activating an assistance feature does not make the vehicle system the legal driver.
Supreme People’s Court Research Office Director Zhou Jiahai said the rules take a firm line on deliberate abuse while calling for caution toward innovation. “The opinion on the one hand emphasizes that criminal acts that deliberately use artificial intelligence to endanger national security, infringe upon citizens’ rights and interests or disrupt social order must be severely punished; on the other hand, it requires that innovative acts in the research and application of artificial intelligence be handled prudently in accordance with the law,” Zhou said.
The court acknowledges that the opinion cannot settle every question. Its own explanation says some issues remain open because courts lack enough experience or consensus to establish a final rule. That includes difficult disputes over model training, the copyright status of AI-generated material, patents involving AI-assisted inventions and responsibility within open-source development chains.
The immediate effect is less dramatic than a new AI law but more concrete for litigation. Courts across China now have a common document telling them how to assess consent, fault, evidence, notice, consumer harm and safety duties in AI cases. The unresolved question is how those general instructions will be applied when the next dispute involves a model whose training data, output and chain of responsibility cannot be cleanly separated.