1. Zhejiang University legal scholars have released a proposed 99-article AI law draft that would require human review of AI-assisted medical decisions, give consumers the right to challenge consequential automated decisions, and impose record-keeping duties on developers and providers [para. 1][para. 2]. The draft, published jointly by several university legal and digital-governance institutions, adds to a growing body of proposed rules as China moves toward national AI legislation, with the Ministry of Justice saying it will accelerate work on an AI law in 2026 [para. 2]. Zhejiang, home to Hangzhou's cluster of AI companies, has become a major testing ground for the technology's commercial use [para. 3].
2. The proposal addresses disputes over whether AI companies may use artists' works to train models, liability when large language models produce false answers, and how people can pursue claims after being targeted by AI-generated face-swapping videos [para. 4]. The eight-chapter document calls for oversight spanning an AI system's full life cycle—from development and deployment to operation and withdrawal—and establishes rules for high-impact uses including health care, justice, education, and finance, while barring applications that manipulate users' emotions or exploit vulnerable groups [para. 5].
3. At the center of the proposal is the "capability-intelligence-task" principle, according to Sun Xiaoxia, head of Zhejiang University's Institute of Digital Rule of Law and leader of the research team [para. 6]. The framework assesses whether a system's operational capabilities, cognitive maturity, and the nature and consequences of a particular task are properly matched [para. 6]. The assessment would help determine whether a system is suitable for a task, conditionally suitable, or unsuitable, providing a basis for tiered regulation, market access, and liability allocation [para. 7]. "The risk of a task depends not only on whether AI is powerful enough, but also on whether it can perform the task reliably and whether the task is appropriate to delegate to it," said Li Wenlong, who led drafting of the proposal [para. 8].
4. The draft rejects a blanket approach to liability for erroneous AI outputs, instead holding developers, service providers, and deployers responsible within the scope of their actual control [para. 10]. Liability would be determined by considering the system's intended use, known limitations, foreseeable risks, and whether necessary safeguards were taken—not judged solely by output accuracy [para. 10]. The draft also states that a company couldn't argue it had met all obligations simply because a human supervisor was nominally involved [para. 11].
5. For copyright disputes over model training, the draft requires developers to establish a lawful basis for use, honor rights holders' restrictions on use, and retain records on data sources and processing [para. 12]. In health care, AI-generated diagnostic conclusions, treatment plans, and medication recommendations could only be used after review by a licensed physician [para. 13]. Financial institutions using AI to make unfavorable decisions must provide a channel for human review with a clear processing deadline, rather than offering only generic explanations [para. 13]. Individuals harmed by AI outputs could file appeals with providers or deployers, which would have 15 working days to process complaints [para. 14].
6. The proposal promotes access to AI development resources, requiring public computing platforms to offer access on fair, reasonable, and nondiscriminatory terms [para. 15]. It backs open-source ecosystems and allows applicants denied access to public computing power, data, platforms, or application scenarios to pursue remedies [para. 15]. The draft prohibits exploiting cognitive characteristics of minors, elderly people, and other vulnerable groups to induce decisions against their interests, and bars targeting people in vulnerable conditions with disadvantageous goods or services [para. 16]. It also prohibits manipulation of consumption decisions through emotional capture, false identities, or subliminal stimuli, giving users rights to information, choice, human review, and appeals [para. 17].
7. The proposal reflects concerns from technology companies about regulatory predictability and compliance costs [para. 18]. Researchers visited more than 20 leading AI and internet companies in Hangzhou while preparing the draft, gathering views on technical limits, deployment conditions, and compliance costs [para. 18]. Companies raised questions about access to computing power and data, public data availability, and whether smaller companies could meet the same compliance demands as large ones [para. 19]. The draft incorporates some of these concerns, calling for compliance flexibility for small and startup companies proportionate to their size and degree of control [para. 20].
8. The university proposal is an academic contribution rather than a legislative bill, but joins a widening policy push in China [para. 21]. Beijing issued rules governing algorithmic recommendations, deep-synthesis technology, and generative-AI services between 2021 and 2025 [para. 21]. China's State Council said in 2017 that the country aimed to establish an initial framework of laws and ethical norms for AI by 2025 [para. 22]. Other Chinese academic groups have also issued model AI laws, including a 2024 scholars' draft with 96 articles and a separate model law led by Chinese Academy of Social Sciences researchers, first released in 2023 and now in its fourth version [para. 23][para. 24].
AI generated, for reference only
Sign in with Google
Sign in with Facebook
Sign in with 财新