On September 7, the Supreme People's Court issued the Opinions on the Lawful Trial of Cases Involving Artificial Intelligence Disputes, document number Fa Fa [2026] No. 10. This is the first time the highest adjudicative body has issued a document specifically on adjudicative rules for AI-related disputes.
First, let us be clear about its weight. It is not a judicial interpretation. The document number is "Fa Fa" rather than "Fa Shi," so judges cannot directly cite it as an adjudicative basis when writing judgments. However, through elevated jurisdiction, case guidance and trial supervision, it unifies the standards of courts at all levels, and its practical impact is comparable to that of a judicial interpretation. Treating it as "a common reference answer for judges nationwide handling AI cases" is not far off.
If you are building AI products, integrating AI into your business processes, or are about to enter the market, a considerable portion of this document is written for you. And its most important change is not how many new obligations it adds. The several measures issued by regulators over the past three years have already set out most of the obligations. What has changed is the nature of these obligations.
The five mechanisms below are not a list directly given in the Opinions; they are the result of consolidating the requirements imposed on service providers across four sets of rules: the Opinions, the Interim Measures for the Administration of Generative Artificial Intelligence Services, the Provisions on the Administration of Deep Synthesis of Internet Information Services, and the Measures for Labeling Artificial Intelligence-Generated Synthetic Content. As for why I suggest acting now rather than waiting for regulators to come knocking or for a notification letter, I first need to explain clearly what the nature of these obligations has actually become.
Compliance mechanisms are no longer just a ledger; they are now your evidence in litigation
In the past, when AI compliance was discussed, the focus was on filing, assessment and inspection. Mechanisms were built mainly to satisfy the competent authorities, made into a ledger that could be produced when they came.
What the Opinions change is this: the same set of mechanisms now directly determines whether you pay damages in civil litigation and how much you pay.
Article (3) first establishes the principle of liability: where the law does not provide otherwise, fault-based liability applies, meaning liability only where there is fault. It then lists a series of factors for determining whether fault exists:
(3) Accurately grasp the principle of liability for AI-related tort liability. Accurately apply the relevant provisions of the Civil Code, the Personal Information Protection Law, the Product Quality Law, the Road Traffic Safety Law and other relevant laws, and determine in accordance with law the legal liability arising from the use of artificial intelligence to infringe civil rights and interests. Where the law does not clearly provide for the application of no-fault liability or presumption-of-fault liability, whether the actor bears tort liability shall be determined in accordance with the fault-liability principle under Article 1165, paragraph 1, of the Civil Code. In determining whether the actor is at fault and the degree of fault, comprehensive consideration shall be given to factors such as the specific scenario of the AI application, the degree of autonomy, the transparency of technology and information, the potential risks and scope of impact, the measures taken and the technical feasibility of measures taken by AI developers, providers and other relevant entities to prevent and reduce AI infringement, and the AI user's ability to foresee and ability to control the damage that may be caused by infringement carried out using AI technology.
In this passage, two items are directly tied to your mechanisms: one is what measures you have taken to prevent infringement and whether they are technically feasible; the other is how transparent your technology and information are. Put plainly, every compliance action you have taken will be converted into a degree of fault in court.
Even more direct is Article (12):
(12) Regulate in accordance with law acts involving artificial intelligence that infringe intellectual property rights. Where AI-generated content infringes another person's copyright, the people's court shall, after comprehensively considering factors such as the type of AI service, industry characteristics, source of training data, degree of participation of the parties, necessary measures taken and profit situation, reasonably determine in accordance with law the liability of the AI developer, provider and user. Where an AI developer raises a non-infringement defense, it shall be ordered to provide evidence such as the source of training data, training process records, model operation mode and scientific theoretical basis. Where a rights holder claims that an AI provider infringes its copyright by using algorithmic technology, it shall provide relevant evidence. Where an AI user knows or should know that a prior work exists, uses AI to generate a work substantially similar to the prior work and has no reasonable defense, and the rights holder of the prior work requests that it bear tort liability, the people's court shall support such request. Where artificial intelligence is used to carry out counterfeiting, false advertising, traffic and order manipulation and other acts that infringe the rights of others or constitute unfair competition, corresponding liability shall be borne in accordance with law.
The sentence in the middle deserves attention: where a developer raises a non-infringement defense, the developer should be ordered to provide the sources of its training data, records of the training process, the model's operating mode, and the scientific and theoretical basis.
Translated, this means: you say I have not infringed. The court will not simply hear that and let it go; rather, it will require you to lay out the evidence. Where the data came from, whether the training process was recorded, how the model runs, and whether it holds up technically. If you can lay it out, the defense stands; if you cannot, the statement "I have not infringed" is just a sentence.
This is somewhat like a dashcam. You install it not for the traffic police to see, but so that if something happens, you can explain for yourself what happened at the time. A compliance mechanism now plays the same role. Of course, the analogy is not entirely precise. A dashcam is only responsible for recording, whereas a compliance mechanism, in addition to leaving records, must itself actually stop infringement; it is both evidence and measure.
The answers in three judgments
This is not speculation. Before the Opinions were issued, the courts had already used three judgments to show the market how this account is settled.
The Guangzhou one, Guangzhou Internet Court (2024) Yue 0192 Min Chu No. 113, judgment pronounced on February 8, 2024. A user entered Ultraman-related prompts on the platform, and the platform output images substantially similar to the Ultraman character. The court found that the platform directly infringed the right of reproduction and the right of adaptation, awarded damages of RMB 10,000, and ordered the platform to adopt keyword filtering measures. In determining whether the platform was at fault, the court identified three matters: lack of a complaint and report mechanism, lack of warnings of potential risks, and lack of prominent labeling.
The Hangzhou one, first instance Hangzhou Internet Court (2024) Zhe 0192 Min Chu No. 1587, second instance Hangzhou Intermediate People's Court (2024) Zhe 01 Min Zhong No. 10332, which upheld the original judgment on December 30, 2024. This platform provided the function of online training of LoRA models; users uploaded Ultraman materials to train models, and other users could overlay this model to stably generate Ultraman images. The court found that the platform did not constitute direct infringement, but did constitute contributory infringement, and awarded damages of RMB 30,000. The platform argued that it merely provided technical services based on open-source models and that the infringing materials were uploaded by users, so it should fall within the safe harbor. The court did not accept this. The platform separately listed well-known IPs, including Ultraman, under LoRA models, directly profited through membership recharge, and could have filtered but did not do so. As for the disclaimer written by the platform in the user agreement, the court's position was blunt: a disclaimer cannot exempt statutory obligations.
The Shanghai one, that is, the Medusa case, first instance judgment pronounced by the Shanghai Jinshan District People's Court on November 3, 2025, second instance Shanghai Intellectual Property Court (2025) Hu 73 Min Zhong No. 1189, which upheld the original judgment on April 29, 2026. It likewise involved users uploading materials to train LoRA models, and likewise generated images substantially similar to prior works. The result was that the user, Li, was ordered to pay RMB 50,000, and the platform bore no liability at all. Where was the difference? The court listed four points: the platform fulfilled its reasonable notification obligation to users, set up a reporting entry on each model page, conducted machine review before publication, and promptly took down all models involved after receiving the complaint and updated its review keywords.
The same type of case, the same technical architecture, and damages ranging from RMB 10,000 to RMB 30,000 to zero. The difference in between is not luck; it is mechanism.
The mechanical gap between the three judgments
The first mechanism is labeling
Many people still understand labeling as "putting a watermark on an AI-generated image." That understanding expired on September 1, 2025.
Article 12 of the Interim Measures for the Administration of Generative Artificial Intelligence Services is itself very short, and it refers the labeling obligation out in its entirety:
Article 12: Providers shall, in accordance with the Provisions on the Administration of Deep Synthesis of Internet Information Services, label generated content such as images and videos.
What catches this referral is Article 17, paragraph 1 of the Provisions on the Administration of Deep Synthesis of Internet Information Services. It delineates which services must carry conspicuous labels:
Article 17: Where a deep synthesis service provider provides any of the following deep synthesis services, which may cause confusion or misrecognition among the public, it shall place a conspicuous label in a reasonable position or area of the generated or edited information content to inform the public of the deep synthesis:
(1) text generation or editing services that simulate natural persons, such as intelligent dialogue and intelligent writing;
(2) speech generation or editing services such as synthesized human voice and voice imitation, or editing services that significantly change personal identity characteristics;
(3) generation of images or videos of persons, such as face generation, face replacement, face manipulation and posture manipulation, or editing services that significantly change personal identity characteristics;
(4) generation or editing services such as immersive simulated scenes;
(5) other services with the function of generating or significantly changing information content.
Note that this provision carries its own limitation: it applies only where the content may cause public confusion or misidentification. Not all AI-generated content must be conspicuously labeled, but intelligent dialogue, voice simulation, and face-swapping are categories that essentially cannot avoid it.
The real upgrade took effect on September 1, 2025. On that date, the Measures for Labeling Artificial Intelligence-Generated Synthetic Content came into force, and the supporting mandatory national standard GB 45438-2025 took effect at the same time. It splits labeling into two layers: explicit labels are what users can see at a glance, such as text prompts, markings on images, and prompts on the opening frame of a video; implicit labels are embedded in file metadata, invisible to the naked eye and readable only with tools.
The legal force of these two layers differs, and this is where mistakes are most likely:
Article 5: Service providers shall, in accordance with Article 16 of the Provisions on the Administration of Deep Synthesis of Internet Information Services, add implicit labels to the file metadata of generated synthetic content. Implicit labels include production element information such as the attribute information of the generated synthetic content, the name or code of the service provider, and the content number.
Service providers are encouraged to add implicit labels in forms such as digital watermarks to generated synthetic content.
Metadata is a "shall" — a hard obligation; digital watermarks are "encouraged" and not mandatory. I have seen quite a few companies invest considerable manpower in watermarks while failing to fill in the metadata fields correctly — precisely the wrong direction. At the same time, do not overestimate metadata. It disappears after screen recording, transcoding, screenshots, or secondary editing, and it is unrealistic to expect it to support long-term infringement tracing.
There is another channel that is easily overlooked — the exception Article 9 leaves for content "without explicit labels":
Article 9: Where a user applies to a service provider for generated synthetic content without explicit labels, the service provider may, after clearly specifying the user's labeling obligations and usage responsibilities through the user agreement, provide generated synthetic content without explicit labels, and shall lawfully retain information on the recipient and other relevant logs for no less than six months.
If your product serves commercial clients who need "clean final cuts," this is your lawful exit, but it is tied to two conditions: the user agreement must clearly state the labeling obligations and usage responsibilities, and logs must be retained for at least six months. Without either one, the exit does not hold.
Looking back at the Guangzhou case, "lack of conspicuous labeling" was one of the three grounds on which the court found the platform at fault — and that was before the labeling measures took effect. That line has now been raised.
The second mechanism is complaints and reports.
Article 15 of the Interim Measures sets this out in considerable detail:
Article 15: Providers shall establish and improve complaint and report mechanisms, set up convenient complaint and report entry points, publish the handling process and feedback time limits, promptly accept and handle public complaints and reports, and provide feedback on the handling results.
Four things are embedded in one article: the entry point must be convenient, the process must be public, the feedback time limit must be published, and the handling results must be fed back. If any one of the four is missing, the mechanism is incomplete.
Article (7) of the Opinions, by contrast, provides a dedicated safe harbor for generative AI:
(7) Prudently determine the tort liability of generative artificial intelligence service providers. Where content automatically generated by generative artificial intelligence infringes upon others' personality rights and interests, such as the right to reputation and the right to privacy, and, after notification by the rights holder, the generative artificial intelligence service provider fails to promptly take necessary measures such as stopping the generation of the infringing content, it shall bear tort liability for the damage caused in accordance with the law. The notification shall include preliminary evidence constituting infringement and the true identity information of the rights holder. Where a network user maliciously induces generative artificial intelligence to generate infringing content by inputting infringing prompts or other means and causes damage to others, that network user shall bear tort liability in accordance with the law; where, after notification by the rights holder, the generative artificial intelligence service provider fails to promptly take necessary measures such as stopping the generation of the infringing content and blocking the relevant generation instructions, and the rights holder requests that the network user and the generative artificial intelligence service provider bear civil liability in accordance with Article 1195 of the Civil Code, the people's court shall support such request in accordance with the law.
This passage has three layers.
The first layer: if the rights holder has notified you and you fail to promptly take necessary measures, you must compensate.
The second layer: the notification itself has a threshold; it must carry preliminary evidence of infringement and the true identity information of the rights holder. A casually sent email does not count. This is protection for the platform.
The third layer differs the most: the necessary measures referred to here are "stopping the generation of infringing content" and "blocking the relevant generation instructions," not deleting a link. Traditional notice-and-takedown targets content already sitting on the server, and deletion ends the matter; generative AI requires you to act on the model side so that it cannot generate the content going forward.
It should be noted that Article (7) governs the personality rights track, such as the right to reputation and the right to privacy. Notice-and-handling for copyright still follows the rules for network service providers in the Civil Code. But the entry point for both tracks is the same: you must first have a place capable of receiving notifications.
The problem in the Guangzhou case lay precisely in the entry point. The platform stated in its user agreement that it does not review content uploaded by users, yet it could not produce evidence proving that it had displayed complaint channels in a prominent position on its website; on that basis, the court found that it had not even fulfilled its basic review and supervision responsibilities. The platform in the Medusa case was the opposite: every model page had a report entry point, and it took content down upon receiving the complaint, ultimately bearing zero liability.
There is another frequently overlooked action called forwarding notice. After receiving a rights holder's notification, you must not only handle it yourself but also forward that notification to the user who posted the content. In the Medusa case, the court found that the platform "objectively fulfilled the obligations to 'take necessary measures' and to 'forward notice'," and forwarding notice was an independent item among them, not incidental.
The third mechanism is record-keeping for training data
This is the most expensive of the five mechanisms, and also the one entrepreneurial teams are most likely to skip.
Article 7 of the Interim Measures sets the baseline for training data:
Article 7 Generative AI service providers (hereinafter referred to as providers) shall lawfully carry out training data processing activities such as pre-training and optimization training, and comply with the following provisions:
(1) Use data and foundation models that have lawful sources;
(2) Where intellectual property is involved, the intellectual property rights lawfully enjoyed by others shall not be infringed;
(3) Where personal information is involved, personal consent shall be obtained or other circumstances prescribed by laws and administrative regulations shall be met;
(4) Take effective measures to improve the quality of training data and enhance the authenticity, accuracy, objectivity and diversity of training data;
(5) Other relevant provisions of laws and administrative regulations such as the Cybersecurity Law of the People's Republic of China, the Data Security Law of the People's Republic of China and the Personal Information Protection Law of the People's Republic of China, and the relevant regulatory requirements of the competent authorities.
Article 19 then sets out the duty to cooperate when regulators come to inspect:
Article 19 Where the relevant competent authorities carry out supervision and inspection of generative AI services in accordance with their duties, providers shall cooperate in accordance with the law, explain the sources, scale, types and labeling rules of training data and the mechanisms of the algorithms as required, and provide necessary technical, data and other support and assistance.
If you place Article 19 and Article (12) of the Opinions side by side, you will find that the materials to be submitted are almost the same document. The only difference is to whom they are submitted: one to the competent authorities, and one to the judge. So record-keeping, once done, can be used twice; this is the only comforting aspect of it.
As for what specifically to keep on record, the national standard Basic Security Requirements for Generative Artificial Intelligence Services sets this out in the most detail. It was initially issued in March 2024 as the technical document TC260-003, and its content was later upgraded to the recommended national standard GB/T 45654-2025. It divides corpora by source into four categories, with different requirements for each: where open-source corpora are used, there must be an open-source license agreement or relevant authorization documents; where data is self-collected, there must be collection records, and data that others have expressly stated may not be collected should not be collected; where data is purchased or obtained through cooperation, there must be a legally effective transaction contract, and where the counterparty cannot even provide commitments as to source, quality and security, the data should not be used; where information input by users is used as corpora, there must be records of user authorization.
This checklist still reads like an administrative requirement, but after the Opinions its purpose has changed. If litigation actually arises and you claim that your model has nothing to do with the other party's work, the judge will not discuss technical principles with you; he will require you to produce these materials.
The fourth mechanism is output-side filtering
Article 14 of the Interim Measures is divided into two paragraphs and governs what to do after a problem is discovered:
Article 14 Where a provider discovers illegal content, it shall promptly take measures such as stopping generation, stopping transmission and elimination, take measures such as model optimization training to carry out rectification, and report to the relevant competent authorities.
Where a provider discovers that a user is using a generative artificial intelligence service to engage in illegal activities, it shall, in accordance with law and the agreement, take measures such as warnings, restricting functions, suspending or terminating the provision of services to that user, preserve relevant records, and report to the relevant competent authorities.
The half-sentence in the first paragraph, "take measures such as model optimization training to carry out rectification," is the key. What it requires is not deletion, but rectification until the content is no longer generated.
The Guangzhou case was the most straightforward. The court not only awarded damages but also ordered the platform to adopt keyword filtering measures and set a verifiable standard: the degree of prevention must be such that when a user normally uses prompts related to Ultraman, images substantially similar to the Ultraman works at issue cannot be generated. This is an outcome standard, not a conduct standard. Whether you have implemented filtering is not what matters; the degree to which you filter is what matters.
The Hangzhou case provided another point of assessment. The court noted that, after superimposing the LoRA model, the platform could stably output specific Ultraman character features, rather than hitting on them randomly. The ability to output stably shows that the matter is technically controllable; if it is technically controllable but not controlled, that is fault.
Following this line of reasoning, there is a quite practical self-test: go through your product features one by one. Which features enable users to stably and repeatedly obtain specific infringing results? Those features are the places where, in the court's eyes, you "could originally have taken measures."
Deleting images is wiping up water on the floor; filtering is turning off the tap. What the court wants is the latter. Of course, the court has not required you to weld the tap shut. The Hangzhou second-instance judgment made it clear that requiring the platform to review every piece of training data entered by users one by one is both technically infeasible and inconsistent with the positioning of a network service provider. The duty of care must match your technical capability and degree of profit. Those who earn more and can control more bear a higher duty.
The fifth mechanism is risk disclosure
The first four mechanisms all cost money. The fifth costs almost nothing, yet it is a type of requirement that appears repeatedly in the Opinions.
Article (9), when addressing AI product liability, places instructions and warnings in the position of a "key review" point:
(9) Accurately determine liability for artificial intelligence products. People's courts shall, in accordance with the definition of "product" in the Product Quality Law, accurately determine in accordance with law artificial intelligence products that take physical objects as their carrier and apply the corresponding legal rules. Where a defect in an artificial intelligence product causes damage, the producer or seller shall bear product liability in accordance with law. In determining whether an artificial intelligence product presents an unreasonable danger endangering personal or property safety, factors such as the nature and use of the artificial intelligence product, its autonomous learning capability, its upgrade and update status, the degree of user control over the system, and whether it conforms to relevant national standards, industry standards and other factors shall be comprehensively considered, and key review shall be conducted as to whether the producer or seller has provided truthful explanations and clear warnings regarding the applicable scenarios, inherent limitations and foreseeable risks of the artificial intelligence product.
Strictly speaking, this article governs AI products with a physical carrier—robots, smart hardware, vehicles and the like—and pure software services do not directly fall within its scope. However, the method of assessment it reveals is general: courts look at whether you have clearly explained "what this thing can do, what it cannot do, and what problems may arise."
What is more telling is Article (13), which leaves a door open for open-source parties, and that door is conditional:
(13) Reasonably determine in accordance with law the legal liability of parties involved in artificial intelligence-related open-source software. In hearing cases involving open-source software, when determining the infringement liability of open-source software developers, providers, and subsequent developers and providers, factors such as the type of open-source license agreement, the specific content of rights restrictions, security and compliance measures, and the degree of information disclosure shall be comprehensively considered, and appropriate liability exemptions shall be granted to open-source software developers and providers in accordance with law. Where an open-source software developer or provider provides, in a free and open-source manner, some code modules required for the research and development of artificial intelligence-related software and publicly explains their functions and security risks, and another person's use of such code modules causes infringement, a people's court may determine that the open-source software developer or provider does not bear infringement liability.
Read carefully: the exemption is not because you open-sourced, but because, while open-sourcing, you "publicly explained its functions and security risks." Explanation is a condition, not a posture. Where the same code modules are released free of charge, the party that wrote a risk explanation may not bear liability, while the party that did not write one falls outside the scope of this exemption.
Among the four grounds on which the platform in the Medusa case was exempted from liability, the first was that it had fulfilled its duty of reasonable notice to users.
This item has the highest cost-effectiveness among the five. Documentation, pop-up windows, risk clauses in user agreements, and explanations of capability boundaries in model cards cost only a few person-days, yet in fault determination it is a genuine factor. Conversely, where a product page boasts that the AI is omnipotent and contains not a single risk warning, then if something goes wrong, those promotional materials will be taken by the other side as evidence.
The five mechanisms, taken together, speak to the same thing
Labeling, complaints and reports, record-keeping of training data, output-side filtering, and risk disclosure—viewed separately, they are five obligations scattered across four sets of rules; viewed together, they answer the same question: after something goes wrong, can you prove that you had already done your best at the time?
The five mechanisms feed into fault determination
Article (3) of the Opinions incorporates "the measures taken and technical feasibility" into the factors for considering fault; Article (12) shifts the burden of proof regarding training data onto developers; Article (18), when reviewing whether AI-generated content can be admitted as evidence, looks at model training, algorithm design, and the content filtering mechanism for generated content. The three articles point to the same place. Judges do not look at what you said, but at what you left behind.
That is the full meaning of the change in the nature of compliance mechanisms. It used to be a cost; it is now an asset, except that this asset only realizes its value on the day you are sued.
If you are in the process of launching an AI product, or have already received a notice letter from a rights holder, and need to conduct a health check against these five mechanisms, you are welcome to talk with me.
Reference Sources
- Opinions of the Supreme People's Court on the Lawful Trial of Cases Involving Artificial Intelligence Disputes (Fa Fa [2026] No. 10, issued September 7, 2026)
- Interim Measures for the Administration of Generative Artificial Intelligence Services (Order No. 15 of the Cyberspace Administration of China and Six Other Departments, effective August 15, 2023)
- Provisions on the Administration of Deep Synthesis of Internet Information Services
- Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content (issued by the Cyberspace Administration of China and Three Other Departments, effective September 1, 2025)
- GB 45438-2025 Cybersecurity Technology — Labeling Methods for Artificial Intelligence-Generated Synthetic Content
- GB/T 45654-2025 Cybersecurity Technology — Basic Security Requirements for Generative Artificial Intelligence Services (formerly the TC260-003 technical document)
- Guangzhou Internet Court, Civil Judgment (2024) Yue 0192 Min Chu No. 113
- Hangzhou Intermediate People's Court, Civil Judgment (2024) Zhe 01 Min Zhong No. 10332; Hangzhou Internet Court, Civil Judgment (2024) Zhe 0192 Min Chu No. 1587
- Shanghai Intellectual Property Court, Civil Judgment (2025) Hu 73 Min Zhong No. 1189
