Artificial intelligence and intellectual property rights: a Chinese perspective

Sylvie Savoie                                                                            

UGGC Avocats, Paris and Shanghai             

s.savoie@uggc.com 

Alison Leroy

UGGC Avocats, Paris

a.leroy@uggc.com

Artificial intelligence (AI) systems are now capable of producing text, images, music, code and technical solutions that rival, and sometimes surpass, human output. This evolution is stress testing traditional intellectual property (IP) frameworks, which were built around human authors and inventors and relatively linear input–output relationships.

China offers a particularly revealing vantage point. It is simultaneously one of the world’s largest digital marketplaces, an AI powerhouse backed by ambitious state strategies and a jurisdiction whose IP system is young, policy driven and closely tied to industrial planning.

As Chinese courts and regulators confront AI, they are being forced to revisit foundational concepts such as ‘authorship’, ‘inventive step’ and the permissible use of data.

This article examines how AI intersects with copyright, patents, training data and trade secrets in China, and how these developments fit into the broader global conversation on AI and IP.

Copyright and the ‘human intervention’ standard

Chinese copyright law, like most copyright systems, is built on a human-centric premise. The Copyright Law protects ‘original intellectual achievements’ created by natural persons.[1] AI systems have no legal personality and cannot be authors in their own right. That simple proposition drives much of the emerging case law.

Shenzhen Tencent v Shanghai Yingxun:[2] AI as a sophisticated tool

In the landmark case Shenzhen Tencent v Shanghai Yingxun (otherwise known as the ‘Dreamwriter’ case), Tencent’s AI system generated an automated financial news article, which was then copied verbatim by a competitor. The Nanshan District Court treated the article as a protected ‘work’ and found that an infringement had occurred.

The reasoning is instructive. The Court focused not on the machine’s autonomy, but on the human input upstream: Tencent’s team had designed the system, chosen the data sources, structured the templates and logic and deployed the tool in a specific editorial context. The employee’s act of triggering the software was framed as a concrete editorial instruction, not a purely mechanical click.

On that basis, the article was recognised as a literary work whose originality stemmed from the human choices embedded in the system’s design and deployment. Authorship and ownership were attributed to Tencent as a ‘legal person work’: the output expressed the company’s will and investment, even though an AI system did the line-by-line drafting.

The message is clear: where humans provide substantial creative arrangement and control, AI is assimilated to an advanced pen. Protection follows the human contribution.

Feilin v Baidu:[3] when human input is too thin

The Beijing Feilin Law Firm v Baidu decision, issued shortly before Dreamwriter, illustrates the other side of the line. Feilin had used Baidu’s ‘Wenxin’/ERNIE platform to generate a legal report and sued a third party for reproducing it.

Here, the Beijing Internet Court found that the report did not qualify as a ‘work’ under the Copyright Law. The user’s role was reduced to entering parameters and clicking ‘generate’, with no demonstrated creative selection, arrangement or editing. In the Court’s view, there was insufficient human intellectual input to ground authorship.

Interestingly, the Court did not leave the plaintiff without any protection: the copying was condemned as unfair competition, characterising the defendant’s conduct as free-riding on another’s investment. But the failure to meet the copyright threshold underscores the emerging spectrum: deep, customised integration and meaningful human control support protection (Dreamwriter), while superficial prompting with minimal intellectual input does not (Feilin).

Between these poles lies a wide grey zone covering AI-assisted workflows in regard to media, design, software and beyond. For now, Chinese courts seem inclined to look for concrete evidence of human intellectual choices, namely in prompts, curation and post-editing, before attaching copyright to AI-related outputs.

Contractual allocation

Even when copyright is arguably present, ownership can be complex. Users, AI service providers and employers may all claim rights. In practice, Chinese actors increasingly rely on the following aspects:

• platform terms of use allocating rights in AI-assisted outputs; and
• internal IP policies clarifying who owns what when employees work with AI tools.

Because statutory rules are still catching up, contract law is doing much of the work in regard to structuring rights in AI-generated content.

Training data, text and data mining and trade secrets

If copyright issues dominate the output side of AI, the input side, the training data, raises its own set of legal and policy dilemmas. Modern AI models feed on massive datasets that mix copyrighted works, personal data and commercially sensitive information. In China, these questions sit at the crossroads of copyright, data protection, trade secrets and economic security.

Copyright and large-scale training

Chinese law does not (yet) provide a broad, explicit exception authorising text and data mining for AI training. The Copyright Law contains traditional limitations and exceptions,[4] but none squarely addresses the legality of large-scale scraping and analysis of protected content to train commercial models.

In practice, Chinese AI developers operate along a continuum, as follows: on one end, licensed corpora, open-source datasets and content created or commissioned for training; and on the other end, more aggressive web scraping of online materials, sometimes justified based on contestable interpretations of existing exceptions or implied licences.

Absent of clear statutory guidance, the legality of training practices tends to be evaluated against general copyright criteria: the substantiality of copying, the nature of the use (commercial vs non-commercial) and the impact on the legitimate interests of rightsholders.

Personal data, ‘important data’ and localisation

Overlaying copyright concerns is a dense layer of data regulation. The Personal Information Protection Law (PIPL), Data Security Law and cybersecurity rules impose stringent conditions on the collection, processing and export of personal information and ‘important data’.

For AI actors, this translates into:

• the need for a lawful basis (consent or otherwise) for large-scale processing of personal data;
• strong security, minimisation and purpose limitation obligations; and
• in many cases, localisation requirements or security assessments for cross-border data transfers.

These constraints have direct IP implications. Data localisation and export controls can fragment training datasets, complicate cross-border research and development (R&D) pipelines and limit the integration of Chinese market data into global models. For multinational companies, aligning AI training practices with both IP and data governance rules is becoming a central compliance challenge.

Models and datasets as trade secrets and the resulting paradox

On the protection side, Chinese law offers robust tools for treating curated datasets, model architectures, parameters and internal tooling as trade secrets under the Anti-Unfair Competition Law[5] and related regulations. For many Chinese AI companies, trade secrets, namely reinforced by non-disclosure agreements (NDAs), access controls and careful internal documentation are the primary IP asset-protecting competitive advantage.

This creates a paradox that is visible globally but particularly acute in China: companies may rely on broad, sometimes opaque access to third-party content for training purposes, while claiming extremely strong secrecy protection for their own trained models and annotated datasets. That asymmetry raises questions of fairness and long-term sustainability. It also fuels policy discussions about whether new, AI-specific regimes, for example, for database-like protection of training corpora or mandatory licensing schemes, may become necessary.

Patents, inventorship and AI-assisted inventions

Beyond creative outputs, AI systems increasingly contribute to technical problem-solving: drug discovery, materials science, engineering design, the optimisation of industrial processes and more. This raises two patent-related questions: can AI be an inventor and how should AI-assisted inventions be evaluated?

No machine inventors: the DABUS experience[6]

Chinese patent law, in line with most major systems, requires that an inventor be a natural person.[7] Patent offices and courts have consistently rejected applications listing an AI system as the inventor.

The attempts by Stephen Thaler to register patents in China (and elsewhere) naming his AI system DABUS as the inventor were summarily refused by the China National Intellectual Property Administration (CNIPA) and the Beijing Intellectual Property Court. The reasoning echoes Dreamwriter: AI is a tool, however sophisticated, not a legal subject. Recognising it as an inventor would contradict the statutory definition of ‘creator’ and unsettle the broader architecture of patent law.

Documenting the human inventive step

The refusal to recognise AI as an inventor does not mean that AI-assisted inventions are unpatentable. Where a human team uses AI to explore solution spaces, simulate scenarios or generate candidates, the resulting technical solution can be patented. But the named inventors must be human, and applicants should be prepared to demonstrate their contribution.

This has both substantive and evidential dimensions. Substantively, patentability still hinges on novelty, an inventive step and industrial applicability, with an emphasis on technical character. Evidentially, applicants are well-advised to document:

• who framed the technical problem;
• who selected and configured the AI approach;
• how human researchers evaluated, selected and refined AI-proposed options; and
• why the claimed solution would not have been obvious to a person skilled in the art.

As AI tools generate thousands of plausible candidates (for example, in molecular design or generative engineering), the ability to articulate and evidence the human ‘inventive step’ becomes critical. Marketing narratives that present the AI system as the true inventor may satisfy consultants and investors, but they risk undermining the legal position before patent offices and courts.

AI-related inventions as patent subject matter

In terms of the subject matter, Chinese practice is broadly convergent with other major offices. AI-related inventions are not excluded per se, but: ‘core AI’ inventions (algorithms, architectures, training methods) must be framed as part of a technical solution, not as abstract mathematical methods and AI-enabled applications (for example, medical diagnosis, autonomous driving, industrial control) are typically assessed at the system level, focusing on the concrete technical effects.

As examination guidelines for AI-related patents evolve, comparative experience, particularly from the European Patent Office, the US Patent and Trademark Office, the Korean Intellectual Property Office and the Japan Patent Office, will likely influence Chinese practice, even as policy priorities (such as promoting domestic AI leadership and managing standard essential patents) give it a distinct flavour.

Conclusion: a policy-driven, evolving framework

From the Chinese perspective, the intersection of AI and IP is currently shaped by three core features. First, there is a firm insistence on human authorship and inventorship: AI systems are treated as sophisticated tools, not as legal subjects, even when they contribute significantly to creative or inventive processes. Second, the regime governing the use of data for AI training is unsettled but tightening. Copyright, personal data protection and economic security rules increasingly constrain how training datasets can be assembled, processed and transferred, especially when personal information or ‘important data’ are involved. Third, companies rely heavily on trade secrets and contractual arrangements to protect AI models, curated datasets and internal workflows, and to allocate rights in regard to AI-assisted outputs.

These legal developments cannot be understood in isolation from the broader policy environment. In China, AI and IP are closely linked to industrial planning, technological self-reliance and data governance. Legislators, regulators and courts are seeking to encourage innovation, secure strategic control over data and digital infrastructure and maintain a degree of regulatory flexibility in the face of rapid technological change. As AI systems become more capable and more deeply embedded in economic life, this framework will continue to evolve. For businesses and practitioners operating in China, the key is to monitor these shifts closely, document human contributions in AI-assisted creations and inventions and align internal policies on data and trade secrets with an IP system that is increasingly designed around AI as both a driver of growth and a strategic asset.