As artificial intelligence models evolve rapidly and their applications broaden, the number of related intellectual property disputes is steadily climbing. Questions such as whether works created using AI can be considered copyrightable, who bears liability when AI-generated content infringes on others' rights, and how to pursue claims in cases involving open-source software are emerging as pressing issues in need of judicial clarification.

On September 7, the Supreme People's Court released the "Opinions on Lawfully Adjudicating Cases Involving Artificial Intelligence Disputes" (referred to as the "Opinions"), which provides systematic rules for handling IP disputes related to AI. On the same day, the "Understanding and Application" of the document, authored by Zhou Jiahai, Director of the SPC Research Office, alongside Deputy Director Si Yanli and other contributors, was made public online.

The Opinions tackle everything from determining liability for AI-generated content that infringes on IP rights, to granting exemptions for developers and providers of free open-source AI software, and to regulating patent authorization and confirmation procedures for AI-related inventions. These measures respond to the challenges that rapid AI technology advancement poses for judicial IP protection. On issues where consensus has yet to be reached, the document reserves judgment, indicating that further experience is needed before clear rules are set out through appropriate channels at a later stage.

Regulating AI-Related IP Infringement

With the rapid iteration of generative AI technology, AI-produced content has penetrated deeply into areas such as content creation, commercial design, and scientific research, presenting new challenges to existing copyright frameworks. Questions about liability for infringement via AI-generated content and the rules for patent authorization involving AI-driven inventions have become hot topics with broad public interest.

The Opinions first make clear that courts must regulate AI-related IP infringement in accordance with the law. When AI-generated content infringes on others' copyrights, courts should reasonably determine the liability of AI developers, providers, and users by weighing factors such as the type of AI service, industry characteristics, data training sources, parties' level of participation, necessary measures taken, and profits gained.

Xiao Yudan, a researcher at the Institutes of Science and Development under the Chinese Academy of Sciences, told Yicai that the Opinions establish, for the first time at the level of the nation's highest judicial body, a tiered liability framework across the industry chain, setting differentiated boundaries of responsibility for developers, providers, and users.

The Opinions also stipulate that if developers raise a non-infringement defense, they must provide evidence such as training data sources, training process logs, model operating modes, and scientific theoretical foundations to support their claim. Xiao highlighted this as the most ground-breaking aspect, as it imposes a mandatory information disclosure obligation on developers for such defenses, effectively enabling judicial "piercing" of the AI "black box."

Regarding patent authorization and confirmation for AI-related inventions, the Opinions specify that such activities must be conducted lawfully. If an AI-related invention adopts technical means that follow natural laws, solves a technical problem, and achieves results consistent with natural laws, courts should recognize it as eligible subject matter for patent law protection, unless it violates the law or social morality, harms public interests, or involves no substantive human contribution. Additionally, if a patent specification's description of the technical solution enables a person skilled in the art to implement the invention, it should be deemed to meet the disclosure sufficiency requirements for authorization.

As for whether AI can be named as an inventor for AI-generated inventions—a contentious issue globally—the Opinions provide clear guidance: where a natural person uses AI to complete an invention and makes a creative contribution to its substantive features, that person should be recognized as the inventor.

Another widely discussed question is whether AI-generated content qualifies as a "work" under copyright law. Yicai noted the Opinions do not explicitly answer this. The accompanying "Understanding and Application" reveals that during drafting, opinions on the copyrightability of AI-generated content were sharply divided, so a definitive rule was withheld.

The document notes that in judicial practice, cases where users of AI services claim copyright over content they generated are not uncommon, with divergent viewpoints emerging. Whether AI-generated content counts as a work under copyright law hinges on whether it meets the statutory requirements. According to Article 3 of the Copyright Law, the three essential elements are: the work is created by human beings, possesses originality, and is expressed in a certain form. However, whether AI-generated content can satisfy these elements remains a major controversy in both theory and practice. Given the lack of consensus—both domestically and globally, where no precedent yet exists for recognizing copyrightability of purely AI-generated content—the Opinions leave this matter undecided for now.

Encouraging Open Source and Clarifying Liability Exemptions

Currently, the development of generative AI models relies heavily on the open-source software ecosystem. A large number of foundational models, deep learning frameworks, training tools, data processing utilities, and middleware are offered under open-source licenses. Handling infringement disputes involving open-source AI is a key practical challenge: how to effectively manage legal risks while protecting and encouraging the enthusiasm for open-source innovation.

The Opinions specify that courts should lawfully determine liability for parties involved in AI-related open-source software. In adjudicating such cases, factors including the type of open-source license agreement, specific content of rights restrictions, security compliance measures, and the level of information disclosure should be comprehensively considered. Courts are also instructed to grant appropriate liability exemptions to open-source developers and providers.

Additionally, if open-source developers or providers offer certain code modules for free, disclosing their functions and security risks to the public for AI software R&D, and downstream users cause infringement through using these modules, the courts may find the developers or providers not liable. The "Understanding and Application" explains that, compared to commercial open-source providers, those who offer free open-source software do not profit from licensing and typically have limited control over how downstream users employ their code. Under principles of balance of rights and responsibilities and matching liability with control, a lower duty of care applies to them, allowing for liability exemptions under certain conditions.

Xiao Yudan believes the Opinions clarify previously ambiguous boundaries of liability for open-source software infringement. By treating license types, rights restriction content, security measures, and disclosure levels as core discretionary factors, the judiciary clearly signals its intent to foster and encourage the open-source ecosystem.

Unresolved Debate: Training AI Models on Copyrighted Works Without Permission

Using publicly available copyrighted works for data training is a widespread practice in AI development. As AI advances, copyright disputes arising from such training have surfaced repeatedly. One notable case involves the New York Times suing OpenAI and Microsoft in U.S. courts for allegedly using millions of its articles to train large models without authorization.

The "Understanding and Application" notes that while this issue was considered during the drafting process, it proved highly contentious. One school of thought argues that training AI models on others' works should fall under fair use, or be deemed non-infringing altogether. The rationale: data training resembles human learning, does not affect normal use of the works, and typically does not reproduce the protected content; moreover, since training involves vast numbers of works with highly dispersed rights holders, requiring individual licenses would impose prohibitive transaction costs and hinder AI development.

The opposing view rejects this stance for several reasons. First, in specific situations, generative AI may reproduce training data or create competing content that harms rights holders' interests. Second, AI data training is complex, with variations in the nature of training data and whether development is commercially motivated affecting the level of protection, making a blanket rule inadvisable. Third, technically, the fair use provisions in copyright law and its regulations do not currently leave room for data training practices.

Due to these deep disagreements, the Opinions leave this issue unresolved for the time being.