(By You Yunting) On September 7, 2026, the Supreme People’s Court of China (“SPC”) released the Opinions of the Supreme People’s Court on the Trial of Cases Involving Artificial Intelligence Disputes (最高人民法院关于依法审理涉人工智能纠纷案件的意见) (“Opinions”). To support the AI industry development, the Opinions introduce a liability exemption for open-source AI software: where a developer publicly discloses the functions and security risks of open-source AI software and a third party’s use of the code module causes infringement, a court may find the developer exempt from infringement liability. However, having reviewed several of the major open-source AI models currently available in China, I found that their disclosures of functions and security risks are generally rather brief and may well fall short of what the exemption requires. This will discuss what an adequate disclosure of the functions and security risks of open-source AI software should look like.
I. Specific Provision on Liability Exemption for Open-Source AI Software Developers
The SPC provides in the Opinions: The legal liability of the parties involved in AI-related open-source software shall be reasonably determined in accordance with law. In adjudicating cases involving open-source software, the court, when determining the infringement liability of open-source software developers and providers as well as that of subsequent developers and providers, shall weigh such factors as the type of open-source license, the specific rights restrictions imposed, security and compliance measures, and the extent of disclosure, and shall, in accordance with law, accord appropriate exemption from liability to open-source software developers and providers. Where an open-source software developer or provider makes available, free of charge and under an open-source license, certain code modules required for the development of AI software and publicly discloses the functions and security risks of such modules, and another person’s use of the code module causes infringement, the people’s court may determine that the developer or provider bears no infringement liability.
II. What the User Agreements of Mainstream Open-Source AI Models Currently Provide
I searched Hugging Face for the licenses governing mainstream Chinese open-source AI models and found that most contain nothing but a bare disclaimer, with no description of functions or security risks at all. DeepSeek-R1 uses the standard English-language MIT License as its user agreement; it runs to only two paragraphs, the second of which is a disclaimer. GLM-5.3’s license has only three clauses, the third being a disclaimer. Kimi-K3’s license has only five clauses, the fifth being a disclaimer. These three disclaimers are broadly the same in substance: the developer makes no warranty as to the software and accepts no responsibility for the consequences of its use.
By contrast, Alibaba’s Qwen3.8-27B does somewhat better. It uses the standard Apache License 2.0, and its disclaimer has three clauses, to the following effect: First, no warranty—the licensor provides no warranties or conditions of any kind, and you must bear the risk of use. Second, limitation of liability—under no circumstances shall any contributor to the open-source software be liable to you for any damages. Third, you may charge a fee, but must bear the resulting liability yourself and agree to release every contributor of the open-source software from liability.
III. Directions for Improvement
The core precondition for exemption under the Opinions is the public disclosure of the software’s functions and security risks; the Opinions further require courts to weigh the specific rights restrictions imposed, security and compliance measures, and the extent of disclosure. At present, however, the terms of mainstream Chinese open-source AI software are nothing more than the traditional blanket disclaimer. Even the relatively comprehensive Apache License 2.0 offers only a passive defense and, measured against the SPC’s requirements, leaves serious gaps:
1. A Disclaimer Is Not Risk Disclosure
In my view, publicly disclosing security risks is an affirmative disclosure obligation—for example, stating expressly that the model may hallucinate, may generate infringing content, and carries an objective risk of being misused for deepfakes. The clauses in those licenses, however, are merely defensive disclaimers. In infringement litigation, a party that has not discharged its affirmative disclosure obligation will find it very difficult to obtain an exemption on the strength of a boilerplate disclaimer alone.
2. The Absence of Functional Descriptions and Defined Technical Boundaries
None of these four licenses describes the functions of the models or code modules, delineates the scenarios in which they may be used (for example, research, commercial, or even military use), or discloses the limitations of the training data. Without such descriptions, a court cannot determine whether a downstream deployer’s use of the module exceeded the module’s reasonably expected scope.
3. No Scenario-Based Restrictions Targeting AI Risks
The specific rights restrictions that the Opinions direct courts to weigh appear in none of the current licenses. These should primarily target high-risk scenarios, such as the generation of unlawful, infringing, or pornographic content, medical diagnosis, autonomous driving, and military use.
In my experience, a license for open-source AI software that satisfies the requirements of the Opinions should at least address the following: the applicable open-source license (e.g., MIT License, Apache License, GPL, LGPL); core functions; intended use; use-case restrictions; known technical limitations; potential security and compliance risks; vulnerability and risk feedback mechanisms; prohibited uses; compliance notices for deployers; redistribution requirements; and disclaimers.
Finally, some readers may ask: training large models burns enormous sums of money; these model developers are willing to spend upwards of RMB 100 million to train a model, so why are their open-source licenses still so thin? In my view, this is partly due to market competition. A deliberately simple license keeps friction in downstream users’ legal review to a minimum, enables the model to be integrated quickly and at scale, and helps win the battle for the developer ecosystem. It is also partly due to industry practice: R&D teams tend to apply open-source standards of the traditional software era wholesale, and no industry consensus has yet emerged on a new type of license specifically addressing open-source AI software risks. Therefore, massive technological investment does not mean mature governance. But now that the SPC has issued these new rules, it is likely that Chinese large-model companies will update their licenses before long to bring open-source AI software into compliance.
Disclaimer of Bridge IP Law Commentary
Short Link: