º¬Ðß²ÝÉçÇø

The rise of DeepSeek has reignited debate over the copyrightability of AI-generated outputs (AI works). Do they qualify as ¡°works¡± under China¡¯s Copyright Law? Can they receive copyright protection? This article examines a judicial ruling to explore the criteria for determining the originality of AI works, offering insights for the industry.

Copyright conundrum

The copyrightability of AI works faces multiple core challenges, spanning the entire process from data training to practical application.

Data training. AI models rely on vast amounts of existing data for learning and reasoning, often generating content that replicates or adapts others¡¯ works. This leads to frequent copyright infringement disputes, exposing companies to legal risks.

Creation. China¡¯s Copyright Law recognises only natural persons and legal entities as authors. This raises questions about the legal status of AI as an author. Do AI works qualify as original ¡°works¡± under the Copyright Law? The industry remains divided on these issues.

Usage. Should the copyright in AI works belong to the AI developer, user, or the public domain? Using others¡¯ works, whether through direct replication or adaptation, and regardless of dissemination, also risks infringing the original rights holder¡¯s copyright.

In copyright disputes, the key issue often centres on whether the work in question qualifies as a ¡°work¡± under the Copyright Law, with the core criterion being whether it meets the standard of ¡°originality¡±.

Judicial perspective

Mainstream legal opinion defines ¡°originality¡± through two key elements: independence and creativity. Independence requires the work to be the result of the creator¡¯s own efforts, ensuring it is distinct from plagiarism. Creativity demands the work reflect the creator¡¯s unique intellectual judgement and choices. Simple physical labour, basic mental tasks or mechanical intellectual outputs fail to meet this standard.

For AI works created through human-machine collaboration, questions arise. Do they satisfy the ¡°independence¡± requirement? To what extent must human intellect be involved to elevate the output beyond mechanical results and demonstrate creativity ¨C 10%, 30% or 70%?

In China¡¯s first copyright infringement case involving AI-generated images, Li Yunkai v Liu Yuanchun (2023), where Li sued Liu for using his AI-generated image without permission, the Beijing Internet Court provided its interpretation of originality as follows.

¡°The plaintiff designed elements such as characters and their presentation using prompts and configured parameters for layout and composition, reflecting his choices and arrangements. After generating an initial image through prompts and parameters, the plaintiff refined it by adding prompts and modifying parameters, ultimately creating the disputed image.

¡°This process demonstrated the plaintiff¡¯s aesthetic preferences and personal judgement. During the trial, the plaintiff generated different images by altering specific prompts or parameters, showing that the model allows users to create varied content through individual inputs.

¡°This indicates that the disputed image is not a mechanical intellectual output. In the absence of contrary evidence, it is concluded that the plaintiff independently created the image, showcasing his personalised expression. Therefore, the image satisfies the criteria for originality.¡±

The case was shortlisted for the Top 10 Nominee Cases Advancing Rule of Law in 2024.

The judgment highlighted the pivotal role of human aesthetic choices. It emphasised that, excluding cases of sole expression where works created through fixed sequences, formulas or structures yield identical results, the copyrightability of AI works hinges on the extent of human intellectual input or contribution.

This approach aligns with traditional copyright standards. Greater human involvement results in richer forms of expression, enhancing the originality of AI outputs. This increases their chances of being recognised as ¡°works¡± and strengthens their position for judicial protection in future disputes.

Suggestions

In this case, the Beijing Internet Court classified AI as a tool, asserting that AI lacks subjectivity and refocusing on the essence of creativity. This aligns with the principles of intellectual property protection and technological innovation. Treating AI as a tool reflects its mainstream market usage, where users generate images, articles, presentations and videos through subjective choices and arrangements. The distinction between products lies in the scope and depth of users¡¯ aesthetic input and the volume of final outputs.

Taking written works as an example, if a user instructs AI to write an article by providing only a title, the level of user involvement is minimal, resulting in low creativity. If the user further refines the output by adjusting paragraph structures or literary techniques, the creative input increases. Greater involvement, such as reconstructing the storyline and character attributes, significantly enhances participation, laying a stronger foundation for originality.

Quantitatively, in the referenced case, the creator selected a model, set six parameters, and used more than 30 positive prompts and 120 negative prompts to produce an artwork that met originality standards. However, this quantitative measure serves only as a reference, and is not a judicial standard.

For companies developing AI-driven products, incorporating refined and specialised human intervention tools during the input stage can enhance human intellectual involvement and produce more diverse content. Additionally, comprehensive AI creation logs should be established to document the complete digital trail from concept to final output.

Liang Qiang is a director and chief partner at Zhong Ce Law Firm. He can be contacted by phone at +86 135 0106 8519 and by email at liangqiang@zhongcelaw.com

Zhang Meijian is an associate at Zhong Ce Law Firm. She can be contacted by phone at +86 185 0042 8480 and by email at zhangmeijianlaw@163.com