AI and copyright: the legal landscape in China

Ping Gu

Zhong Lun Law Firm, Beijing

guping@zhonglun.com

China has rapidly emerged as one of the world’s most active jurisdictions regarding artificial intelligence (AI)-related copyright litigation and regulatory oversight. As generative AI reshapes creative industries, Chinese courts and regulators are working to balance innovation with the protection of intellectual property rights.

The regulatory architecture and legal framework

Below are several pivotal laws and regulations that form the core of China’s AI copyright framework.

The Copyright Law of the People’s Republic of China (2020 Revision)

Article 3 defines the elements of a ‘work’ as ‘original intellectual achievements in the fields of literature, art and science that can be expressed in a certain form’. Article 10 enumerates exclusive rights including reproduction and communication through information networks. Article 24 provides an exhaustive list of 13 fair use scenarios.[1] Whether the reproduction of copyrighted works during AI training constitutes infringement and whether the fair use exception applies must be assessed using this framework.

Regulation for the implementation of the Copyright Law of the People’s Republic of China (2013)

Article 2 defines a ‘work’ as ‘an original intellectual creation in the literary, artistic or scientific domain, capable of being reproduced in a tangible form’. Article 3 states that ‘creation’ means ‘intellectual activity that directly produces literary, artistic or scientific works’.[2]

These provisions constitute the core legal grounds for assessing whether AI-generated content qualifies as a copyrightable work. Specifically, they govern two key factual inquiries: how ‘originality’ is embodied within such content and whether the human intellectual contributions meet the legal threshold for the direct creation of expressive content.

The interim measures for the administration of generative AI services (2023)

Article 7 requires generative AI service providers to ‘use data and base models with lawful sources’ and to ‘refrain from infringing intellectual property rights enjoyed by others’.[3]

This rule establishes a fundamental compliance requirement for the legitimacy of AI training data. However, it neither explicitly clarifies whether deploying copyrighted works for AI model training falls within the scope of fair use nor does it set forth a standalone statutory exception specifically for text and data mining (TDM) activities.

Opinions of the Communist Party of China Central Committee and the State Council on Establishing a Data Base System to Maximise a Better Role of Data Elements (the ‘Data Twenty Articles’, 2022)

This policy document puts forward an innovative data property rights framework built upon a tripartite separation of rights: the right to possess data resources; the right to process and utilise data; and the right to operate data products. It offers fundamental guidance for clarifying the ownership of AI training data and the rights attribution of training datasets categorised as data products.[4]

Cybersecurity technology: the labelling method for content generated by AI (GB 45438-2025)

Released on 28 February 2025 and taking effect on 1 September 2025, this national mandatory standard specifies detailed rules for explicit and implicit content labelling, applying to both AI content generation and distribution service providers. Combined with the labelling measures, it forms a complete operational framework for AI content traceability.[5]

Cybersecurity technology: basic safety requirements for generative AI services (GB/T 45654-2025)

Released on 25 April 2025 and taking effect on 1 November 2025, this standard lays down the requirements for training data security, model safety and the corresponding security safeguards. As a technical standard that provides support to the Interim Measures, it centres on risk identification for training data, the reliability of model outputs and adaptability across service scenarios.[6]

The Regulations on Network Data Security Management (State Council Order No. 790, effective 1 January 2025)

Article 19 explicitly requires network data handlers providing generative AI services to strengthen security management in relation to training data and any related processing activities. The Regulations further mandate that automated tools for network data access and collection must evaluate their impact on network services, refrain from unauthorised network intrusion and avoid disrupting regular network operations.[7] These provisions establish clear compliance boundaries for data scraping activities carried out during AI model training.

The ‘Qinglang Rectifying AI Application Malpractice’ special campaign (2026)

Launched by the Cyberspace Administration of China on 30 April 2026, this four-month campaign targets the legality of training data sources and the unauthorised exploitation of text, image, audio and video data for model training. It also addresses failures to comply with mandatory labelling rules for AI-generated content, reflecting the move to closer regulatory scrutiny of both AI training data and content-labelling obligations.[8]

Copyright issues involving AI training data

The core controversy turns on whether the automated crawling, reproduction and analytical processing of copyrighted literary, artistic and audiovisual works for large-model training constitutes copyright infringement and whether such activities qualify for an exemption under the statutory fair use provisions contained within Chinese Copyright Law. Unlike jurisdictions with an explicit carve-out for TDM, China’s current legal regime lacks targeted legislative provisions governing AI training practices, leaving dispute resolution heavily reliant on judicial interpretation and case-specific adjudication. Meanwhile, intensifying regulatory scrutiny of training data compliance, as exemplified by the 2026 Qinglang special campaign,[9] has significantly raised the level of legal risk associated with unregulated data collection and utilisation for AI developers.

As a landmark ruling concerning generative AI and copyright, the ‘Medusa’ AI large-model copyright infringement case sets clear boundaries for the lawful use of training data and establishes critical compliance benchmarks for the AI industry.[10] The defendant AI service provider crawled and reproduced the plaintiff’s copyrighted online literary works without authorisation to compile training datasets for its large language model. The defendant contended that algorithm training, which was non-public and technical in nature, constituted fair use under the Copyright Law.

The court rejected this fair use defence and held that:

• the training activities substantially enhanced the competitiveness and commercial value of the defendant’s AI services;
• the literary works at issue were entitled to full copyright protection and carried obvious commercial value; and
• the unauthorised bulk exploitation undermined the plaintiff’s rights to exploit its works commercially and disrupted the normal order of copyright licensing.

The court, therefore, held that the defendant’s use of unlicensed copyrighted works for AI model training constituted copyright infringement. The defendant was ordered to discontinue the use of the infringing training data and pay damages. This ruling explicitly confirms that commercial AI model training may not rely on the fair use doctrine without explicit authorisation from the relevant copyright owners.

Beyond this landmark case, China’s judicial and regulatory authorities have established a consistent approach to governing AI training data. For non-public, non-profit AI training conducted for scientific research purposes, courts and regulators take a permissive stance and acknowledge the applicability of fair use. In contrast, any unauthorised scraping, reproduction and exploitation of copyrighted works for commercial generative AI model training will be deemed copyright infringement.

Furthermore, in accordance with the requirements set out in the Interim Measures for the Administration of Generative Artificial Intelligence Services,[11] AI service providers are required to verify the lawful provenance of all of its training data, remove infringing content from datasets and establish standardised data review and management mechanisms. Lawfully sourced training data is, therefore, a basic requirement for generative AI compliance.

The eligibility of AI-generated content for copyright protection

The legal standard: substantive human intellectual contribution

The Copyright Law lays the doctrinal foundation for resolving authorship, ownership and copyright infringement disputes over protected works. Article 3 defines a ‘work’ as ‘an original intellectual creation in the fields of literature, art and science that is capable of being fixed in a certain form.’[12] The threshold issue in AI copyright disputes is whether AI-generated outputs satisfy this statutory definition. This inquiry hinges on whether a natural person has contributed sufficient original intellectual effort to the expressive elements of the resulting work.

Chinese jurisprudence has consistently upheld this human-centric standard. Courts have articulated the governing test as follows: content created primarily through automatic AI generation shall not be deemed a copyrightable work. Yet where a user uses AI merely as a tool and the output reflects the user’s original intellectual creativity, the resulting work may qualify for protection under the Copyright Law.

In its Implementation Plan for Judicial Protection of Intellectual Property Rights (2026–2030), the Supreme People’s Court has instructed lower courts to assess two key aspects when judging whether AI-generated content embodies a natural person’s original selection and expression: the specifics of human input prompts, as well as the actual selection and revision process involved.[13]

Evidentiary framework: burden of proof and process documentation

A party claiming copyright in AI-generated works bears the affirmative burden of proving their creative contribution. In Li v Liu, the Beijing Internet Court found that post hoc reconstructions of the generation process lack adequate probative value.[14]

The plaintiff Li used the open-source AI image model Stable Diffusion to generate a distinctive AI artwork by crafting detailed prompts, adjusting model parameters and conducting iterative selection. After Li posted the image online and claimed exclusive copyright in it, the defendant Liu reposted it without authorisation and contended that such content lacked human creative authorship and was not copyrightable.

The Court rejected this defence, holding as follows:

• the plaintiff’s intentional prompt engineering, parameter tuning and repeated selection reflected personalised aesthetic judgment and original intellectual creation;
• the final image possessed unique expressive characteristics and satisfied the statutory ‘originality’ threshold for copyrightable works; and
• the unauthorised republication infringed the plaintiff’s right of attribution and information network transmission.

Accordingly, the court affirmed the copyrightability of the AI-generated image and awarded damages.

This ruling carries important practical implications: human guidance, personalised selection and iterative optimisation during AI generated content creation qualify as valid creative inputs. Creators and businesses must, therefore, systematically preserve prompt records, parameter settings and iteration histories to prove originality and obtain copyright protection for AI works.

Complementary protections: trade secrets and AI

The Provisions on the Protection of Trade Secrets (State Administration for Market Regulation Order No. 126, effective 1 June 2026) have direct implications for AI development and deployment. Article 5 explicitly lists data, algorithms, computer programs and code as protectable information. Article 7 extends protection to failed experimental data and unsuccessful technical solutions. Article 9 clarifies that reasonable confidentiality measures, applicable to remote work, cross-border collaboration and data processing, include tiered access controls, data anonymisation and the retention of operation logs, establishing a three-pronged framework of contractual, managerial and technical safeguards.[15]

The Provisions also allow Chinese regulators to investigate and take action against overseas infringements that undermine market competition in China. They also require security assessments for cross-border transfers of confidential information and equivalent confidentiality safeguards by overseas recipients, protecting AI-related trade secrets in cross-border operations.[16]

Trade secrets law complements copyright in protecting AI assets. Copyright safeguards the expressive form of original works, while trade secrets law may protect the underlying data, algorithms, model parameters and development processes.

Conclusion

China’s AI copyright legal framework has evolved rapidly from a reliance on general principles to a sophisticated ecosystem comprising specialised regulations, technical standards and judicial precedents.

Several key trends merit close attention. The human‑centred originality standard is expected to endure, further strengthened by the Supreme People’s Court’s judicial implementation plan. For enterprises, it will be critical to actively monitor this fast-evolving regulatory landscape, maintain comprehensive documentation of AI training and generation workflows and implement robust trade secret protections for model assets. The Provisions on the Protection of Trade Secrets are likely to trigger a new wave of litigation and enforcement action, further underscoring the imperative of proactive compliance.