Supreme People's Court AI Judicial Opinion Takes Effect: Enterprise AI Assets and Liability Boundaries, All 24 Rules Explained at Once
On September 7, 2026, the Supreme People's Court issued the *Opinions on Lawfully Adjudicating Cases Involving Artificial Intelligence Disputes* (Fa Fa [2026] No. 10, comprising 5 parts and 24 articles). This is China's first set of adjudication rules specifically governing AI-related disputes.
Bottom line in one sentence: The problem of AI companies previously having to "guess at boundaries" now has official judicial standards — but most business leaders are still unaware of whose side the standards lean toward, or what actions companies need to take.
Below is a breakdown of all 24 key points that every company's AI business should benchmark against.
I. Three Overarching Principles to Remember First
Fault-based liability is the default. Where the law has no special provisions, AI torts are governed by the fault-based liability rule under Article 1165 of the Civil Code; generative AI torts are expressly excluded from strict liability. This is favorable for the industry: it is not a case of "anyone can sue you into the ground whenever AI causes harm."
Liability is tiered by degree of control. The duty of care and fault determination for developers, providers, and users are assessed separately: the higher the degree of autonomy, the heavier the developer's duty of care; the user's ability to foresee and control harm determines their fault.
Innovation is supported, but red lines must be observed. Three principles: human-centered approach, support for innovative development, and a firm bottom line on safety.
II. Four Scenarios Most Likely to Trigger Liability
1. AI face-swapping and voice cloning, AI resurrection of the deceased (Article 4). Generating identifiable virtual likenesses of natural persons without consent and using or publicly disseminating them infringes the right to name and right to portrait; using another person's voice as training data to generate identifiable synthetic voices infringes voice-related rights and interests; manipulating virtual likenesses to publish false statements infringes the right to reputation. E-commerce livestreaming, digital human marketing, and film/visual effects industries should particularly audit their authorization chains.
2. Online doxxing and cyber manhunts (Article 5). Using AI to track and analyze publicly available information for the purpose of prying into privacy, obtaining private information, or disturbing one's peace of life constitutes privacy infringement. Market research and anonymized aggregate analysis fall outside this scope — "with the purpose of prying into privacy" is the key dividing line.
3. Big-data price discrimination and celebrity-likeness endorsement for product sales (Article 10). Algorithmic differential pricing that substantially restricts consumers' right to know, right to choose, or right to fair dealing constitutes infringement; where celebrity-likeness endorsement for product sales constitutes fraud, the punitive damages provisions of the Consumer Protection Law apply (refund of purchase price plus triple damages).
4. Autonomous driving and assisted driving (Article 11). Liability is still determined under the current Road Traffic Safety Law and product liability rules: where harm is caused by vehicle defects, claims may be brought against the manufacturer or seller; where defects are combined with driver fault, both may be sued jointly; automakers bear liability for false advertising regarding "autonomous driving." The mandatory national standard for assisted driving, GB 47955-2026, was issued in July 2026; after the standard takes effect, defects will be determined in accordance with the standard.
III. For Those Engaged in AI Training and Product Development, These Four Provisions Are Key
Use of publicly available personal information in training data (Article 6). Processing information that individuals have self-disclosed or that has been lawfully made public, within a reasonable scope and where the individual has not expressly objected, generally does not constitute infringement; where the processing has a material impact on individual rights and interests, consent must be obtained. "Reasonable scope" is assessed by three factors: the necessity and appropriateness of the processing purpose relative to the model's functions, the type and sensitivity of the information, and the context of public disclosure and reasonable expectations. Note: where a platform has declared that "content will not be used for training" but uses it without authorization, such use exceeds the permitted scope.
Generative AI providers may invoke a safe harbor by analogy (Article 7). Where a rights holder gives notice and the provider fails to take necessary measures in a timely manner — such as ceasing generation or blocking prompts — the provider bears liability for the expanded harm; users who maliciously induce the generation of infringing content bear liability themselves. The red flag rule is not written into the Opinion, but where the provider knew or should have known, Article 1197 of the Civil Code may be applied by analogy.
AI-generated content infringing copyright (Article 12). Where a user knows or should know of the existence of a prior work and still generates substantially similar content, this constitutes infringement, and "it was generated by AI" is not a defense; where a developer raises a non-infringement defense, it must self-certify the sources of its training data, the training process, and the operational model. When a rights holder brings suit, the burden of proof is allocated as follows: to sue the developer, one must prove "generation + substantial similarity"; to sue the user, one must prove "access possibility."
AI inventions and open-source code (Articles 13 and 14). Providers of code modules that are free and open-source with publicly disclosed functionality and security risks may be granted liability exemption (commercial open-sourcing does not qualify for exemption). Where a natural person uses AI to complete an invention-creation and makes a creative contribution to its substantive features, that person shall be recognized as the inventor — the window for algorithmic patent portfolio development is now open.
IV. Data Rights and Interests (Article 16): Dual-Track Protection Under Trade Secret Law and the Anti-Unfair Competition Law
Data rights and interests lawfully obtained are protected; data constituting compilation works or trade secrets is protected under the Copyright Law and the Anti-Unfair Competition Law, respectively; data not constituting trade secrets that is obtained or used by improper means is governed by Article 13, Paragraph 3 of the Anti-Unfair Competition Law (the new data-specific provision added by the 2025 amendment). Fabricating or interfering with data, malicious labeling, and adversarial sample attacks that impair AI operational safety give rise to liability.
V. New Litigation Rules — A Must-Read for Legal Counsel (Articles 17–19)
Adverse inference for evidence obstruction: Where the party controlling evidence refuses to produce it without justifiable cause, the court may deem the opposing party's claims established — AI companies sued should not hide data. Where AI-generated content is submitted as evidence, courts will examine prompt design, similarity, and consistency across repeated testing. The most significant provision: litigation participants submitting AI-generated litigation documents or case retrieval reports must verify their authenticity before the hearing, disclose the use of AI assistance in court, and bear responsibility for the truthfulness and accuracy of the content — lawyers have already been criticized and educated by courts for submitting AI-fabricated case precedents.
VI. Two "Reserved" Issues — Don't Be Misled
Whether AI-generated content constitutes a "work" (copyrightability) and the characterization of using others' works to train large models without permission are deliberately left unaddressed by the Opinion — the controversies are too significant, and resolution is left to case-by-case accumulation. When discussing business externally, remember: there is currently no official conclusion on these two points.
Three Things Companies Should Do Now
First, create audit trails of training data sources, training processes, and operational mechanisms. The Opinion has placed the burden of "self-certification of innocence" on developers — if evidence is not preserved now, there will be no basis for defense in future litigation.
Second, conduct compliance reviews of AI products: authorization chains (portrait and voice materials), notice-and-takedown mechanisms, AI use disclosures in user agreements, and review of e-commerce livestreaming scripts.
Third, subject litigation materials to a "manual verification checkpoint": all case precedents, legal provisions, and evidence should be manually re-verified before submission to the court.
Our two firms have converted this Opinion into a corporate compliance checklist: Aipunajie handles algorithmic patent portfolio development and AI inventor applications, while Mili handles AI infringement and data dispute litigation. Consultation is welcome as needed: 010-65150974.
Further reading: [AI Model Copied — Three Paths to Choose: One Won 1.6 Million, Another Lost 5.1 Billion](/articles/20260907-ai-model-asset-three-track-decision.html)
*This article represents only the author's personal views and does not constitute legal advice. For case-specific analysis, please follow the WeChat public account "Nanjie Mili".*
He Zigang | Intellectual Property Lawyer | Aipunajie · Mili · Najie