Featuring three essential Q&As on AI model protection, liability for AI-generated content, and compliance in overseas markets.

In the digital age, companies are increasingly leveraging Generative AI to boost competitiveness and drive innovation. While this technology enhances content creation efficiency, reduces operational costs, and transforms marketing strategies, it also introduces significant legal risks. As countries around the world continue to explore, implement and reform legislation on artificial intelligence, regulatory frameworks are rapidly evolving.

This chapter focuses on key IP issues in the context of Generative AI, including the protection of models, algorithms, and training data; the infringement risks related to AI-generated content; and compliance considerations for deploying AI models in overseas markets. We have selected three key Q&As from Chapter 1: Generative AI and Intellectual Property Protection of the Practical Q&A Guide to Cutting-Edge Intellectual Property Issues, co-authored by Wolters Kluwer, Rouse, and its strategic partner Lusheng Law Firm, to share.

Selected Practical Q&As

【Intellectual Property Rights of AI Models】Can generative AI models/algorithms and training data be protected by intellectual property rights?

The term “generative AI model/algorithm” in this question refers to a model/algorithm that includes the processes of data training and content generation, has the ability to analyze input data, and can generate content based on the input data.

The protection of generative AI models/algorithms and training data that are based on intellectual property can be considered from the following perspectives:

• Copyright: Generative AI models/algorithms may contain source code and algorithms. These elements themselves may constitute computer software. On this basis, generative AI models/algorithms can be protected by copyright law. The training data may also be protected by copyright law if it contains copyrighted works or constitutes an adaptation or compiled work.

• Patents: If a generative AI model has a unique technology, method or application – such as an innovation used in the model’s training – then the model can also be considered for patent protection as a technical solution. For AI models/algorithms that are more susceptible to reverse engineering, patent protection may be a more suitable choice. From this perspective – where disclosure is exchanged for protection – it is wiser to choose patent protection than trade secret protection. However, it is important to note that patent protection may be challenging, as proving the innovativeness and non-obviousness of AI models/algorithms is not always straightforward.
• Trade secrets: Trade secrets refer to technical/business information that is not known to the public, has commercial value, and has been kept confidential by the right holder. Thus, content with commercial value that is not yet known to the public can theoretically be protected as a trade secret, as long as confidentiality measures are in place.

The advantage of trade secret protection, as compared to copyright and patent protection, is that as long as the AI model/algorithm is still kept secret, the protection can be continued. AI models/algorithms have commercial value and can gain advantages and benefits for the developer. In particular, trade secret protection is a more sensible choice for AI models/algorithms that are not easily obtained by reverse engineering, and for AI models/algorithms whose value lasts longer than the term of copyright and patent protection.

• Property rights and interests under the Anti-Unfair Competition Law and the Civil Law: As noted above, AI models and algorithms are the result of significant investment, requiring human, material and financial resources, along with processes such as program writing, data collection and collation. Additionally, these models and algorithms hold commercial value, providing competitive advantages and economic benefits to developers. Therefore, they may be protected as property interests under the Anti-Unfair Competition Law and Civil Law.

Regarding the training data, large datasets that incur costs and possess commercial value have in some cases been recognized as independent property interests.

Legal precedents suggest that when big data meets these two characteristics – i.e. economic investment and commercial value – then that data product should enjoy independent property interests.[1]

Additionally, when evaluating the IP protection of generative AI models / algorithms, it is necessary to consider the distinction and boundaries from existing algorithms in the public domain.

【AIGC infringement】If the AIGC is similar to an existing work, does it constitute infringement?

The core of copyright infringement in the context of generative AI is whether the prior work has been used without authorization during the AI learning process. If the AIGC is identical or substantially similar to the prior work of another person, there is a higher risk of copyright infringement if it can be proved or presumed that the prior work was used without authorization during the AI learning process. However, if the right holder cannot prove that the prior work was “fed” to the AI model during the AI learning process, or the developer of the AI model can prove that the AI model has never been in contact with the prior work, then the risk of copyright infringement is limited.

Specifically, AI learning and AI creation are two independent processes, which should be evaluated separately:

• From the perspective of AI learning, the generation of AI models inevitably involves learning from data, among which there may be materials that constitute copyrighted works. The act of collecting copyrighted materials for “feeding” constitutes a reproduction act under the Copyright Law, and using the copyrighted works in the process of AI learning and training will constitute use of the works. If the above-mentioned acts are conducted without right holders’ permission, copyright infringement might be established. Currently, the fair use doctrine listed in China’s Copyright Law does not include the reproduction and use of copyrighted works in AI learning and training. Thus it is still inconclusive whether fair use can be invoked as a defence.
• From the perspective of AI creation, considering the process as different from the human creation process – and that there is a “black box” between the AI input data and the output results (which causes a certain degree of uncertainty and uncontrollability between input and the output) – it is not easy to claim and further prove that copyright infringement of reproduction and plagiarism exist during the AI creation process. Nevertheless, courts may still recognize that the AI model is infringing on copyright, based on the presumption of infringing reproduction and the use of copyrighted works during the AI learning process. Specifically, infringement may involve infringement of the right of reproduction, the right of communication through an information network, the right of adaptation, etc. This is especially so when the prior works involved in the above-mentioned scenarios are works with a high degree of originality, or are well-known works that have already been published.

In the “Ultraman” case^[2] heard by the Guangzhou Internet Court, the court adopted the general principle of “substantial similarity + probability of access” when judging the copyright infringement. The court held that the AI-generated part of the involved picture was substantially similar to the original expression of the image “Ultraman”. The court also held that the work of Ultraman enjoyed a high reputation. Thus, it was presumed that the defendant had the possibility of accessing the involved work, and the defendant had constituted copyright infringement of reproduction. Furthermore, the court held that the involved AI generated picture partially retained the original expressions of the work “Ultraman Tiga Multi Type”. On the basis of retaining these original creative expressions, new features were formed, which constituted an adaptation of the plaintiff’s work and thus infringed upon the plaintiff’s right of adaptation regarding the Ultraman work.

As mentioned above, the current judicial practice of the fair use doctrine in China’s Copyright Law is still based on the behaviours of fair use/restriction explicitly listed in the legal provisions, which do not include the reproduction and use of copyrighted works in AI learning and training. Until the fair use doctrine in China’s Copyright Law is amended to embrace the development of AI technology, it is difficult to eliminate the risk of infringement in the data training stage.

However, to encourage the development of AI technology, scholars have begun discussing where AI training data should be included within fair use. Government departments and authoritative institutions have also successively issued AI-related incentive policies and guidance. As China attaches increasing importance to the development of AI technology, a more flexible legal approach may emerge in the future. For example, AI learning and training may be incorporated into “fair use” or certain exemptions may be granted in other ways.

In summary, the issue of copyright infringement of AIGC is currently a cutting-edge, controversial and unclear issue. It is recommended to actively follow the development of legislation and judicial practice, and to explore the boundaries of fair use in AI creation.

In addition, it is also necessary to pay attention to the risks of unfair competition that may be caused by AI creations and AIGC (market confusion under Article 6 of the Anti-Unfair Competition Law, including infringement of merchandising rights, etc., or the good faith principle under Article 2 of the Anti-Unfair Competition Law), as well as  the risk of infringement of portrait rights.

【Generative AI Going Overseas】What IP compliance issues may generative AI model products face when they go overseas?

The most important aspect for generative AI model products going overseas is paying attention to local laws and regulations on AI model product supervision, data training and other issues. In terms of intellectual property compliance, the following points can be used for reference:

• Data compliance. On one hand, it is essential that training data is properly authorized. This can be achieved through commercial agreements with major publishers, copyright holders and other third parties to enhance compliance, particularly concerning intellectual property rights.

On the other hand, legal exemptions for data training and machine learning remain limited. While some countries, such as Singapore and Japan, have introduced specific exemptions, many regions have yet to establish clear policies. Even in jurisdictions where policies exist – such as the EU – significant challenges remain in implementation. Therefore, it is crucial to closely monitor legislative and judicial developments worldwide and to adapt data compliance strategies accordingly.

• Disclosure Obligations. Some countries have already introduced disclosure requirements for AI model training data. For instance, in the EU Artificial Intelligence Act, it is proposed that developers should draft and publish comprehensive descriptions of the data used to train general-purpose models, including copyright data.
• Open-Source Obligations. Some countries have already set regulations addressing open-source issues related to AI models. For example, the EU Artificial Intelligence Act requires that for general-purpose AI models, extensive documentation and guidelines on the use of the model should be made available to downstream providers.

More specifically, for generative AI models expanding to international markets, it is crucial to assess whether the model’s algorithms, software and programs involve open-source components. It is also necessary to consider whether the code generated by the AI model is subject to open-source obligations, and whether it is reflected in the user agreement^[3].

Other Notable Q&As from This Chapter

• 【Definition】What is generative AI?
• 【Copyright of AIGC】Can content such as images, music, and text output by generative AI be protected by copyright? 
• 【Right Holder of AIGC】If the generative AI content output can be protected by copyright, who is the right holder?
• 【AIGC infringement】Does the content output of generative AI constitute an infringement when it is highly similar in style to another person’s prior work?   
• 【Data Training】What authorizations should be obtained for the materials used in the training process of generative AI models?
• 【Liabilities of Generative AI Service Providers】If the AIGC created by users using Generative AI services constitutes infringement, should the AI service providers be liable?    
• 【Responsibility of Platforms Providing GAI Services】How should platforms providing generative AI services deal with an infringing notice that claims the AIGC is infringing?

References

1. Hangzhou Railway Transport Court (2017) Zhe 8601 Min Chu No. 4034 Civil Judgment; Hangzhou Intermediate People's Court (2018) Zhe 01 Min Zhong No. 7312 Civil Judgment.

2. 2024 Yue 0192 Min Chu No. 113

3. WIPO - Generative AI: Navigating Intellectual Property

Chapter Contributors

Landy Jiang, Managing Partner, Global Co-Deputy Head of Dispute Resolution, Lusheng Law Firm, ljiang@lushenglawyers.com

Jenney Zhang, Associate, Lusheng Law Firm, jzhang@lushenglawyers.com

Request a Full Copy

In collaboration with our strategic partner Lusheng in China and Wolters Kluwer, Rouse has developed a valuable resource for rightsholders: “The Practical Q&A Guide to Cutting-Edge Intellectual Property Issues in China”. This guide, compiled by over 30 senior China IP experts from the two leading IP firms, addresses the key concerns of businesses by providing insights on patents, trade marks, copyright, trade secrets, internet unfair competition, intellectual property investment, and punitive damages in an accessible Q&A format. It offers readers the latest legal interpretations, case studies, and practical guidance applicable to their operations.

To request a full copy, please complete the form through the link here.

Please note due to publishing restrictions we may not be able to fulfil every request. The application will be reviewed, and the report will only be available to corporate organizations. Thank you for your understanding

TAGS

• Generative AI
• infringement
• Compliance