AI & Intellectual Property 2026 Comparisons

China does not have a standalone, comprehensive AI statute, but AI-related activities are governed by, among other sources, the following significant legislation, regulations, and official guidance:

• the Cybersecurity Law (effective 1 January 2026);
• the Interim Measures for the Management of Generative AI Services (effective 15 August 2023) (the “Interim Measures”);
• the Administrative Provisions on Deep Synthesis in Internet-based Information Services (effective 10 January 2023);
• the Administrative Provisions on Recommendation Algorithms in Internet-based Information Services (effective 1 March 2022);
• the Measures for Labelling AI-Generated and Synthesized Content (effective 1 September 2025); and
• the Interim Measures for the Administration of Anthropomorphized Interactive Artificial Intelligence Services (effective 15 July 2026) (the “Interactive AI Measures”).

Key Differences

The following are some key differences compared with generally applicable IP law.

The copyrightability threshold

A landmark case issued by the Beijing Internet Court held that sufficient prompt engineering – the selection, arrangement and iteration of prompts – can constitute sufficient “intellectual human input” for copyright protection.

No TDM exception

China’s Copyright Law contains no text and data mining (TDM) safe harbour, making training on copyrighted works legally riskier.

Transparency and pre-market obligations

General IP law imposes no labelling, data source disclosure, or pre-market approval requirements. The PRC framework mandates labelling of AI-generated content and the use of lawfully sourced training data. During regulatory inspections, providers must disclose their training-data sources to the competent authorities. Generative AI services with public-opinion attributes or social-mobilisation capacity must also complete algorithm filing and security assessments before launch. These obligations have no parallel in conventional IP law and reflect China’s hybrid approach treating AI governance as both an IP and a state security/content regulation matter.

China adheres to the Berne Convention and the TRIPS Agreement, ensuring the principle of “National Treatment” for foreign rights-holders, though these foundational instruments predate generative AI and contain no express provisions addressing AI-generated content. At the regional level, the Regional Comprehensive Economic Partnership (RCEP) among 15 Asia-Pacific economies requires effective action against IP infringement in the digital environment. At the policy level, China has actively promoted international co-operation on AI governance, including through the establishment of the World Artificial Intelligence Cooperation Organization (WAICO). China has also issued an AI Co-Operation Development Action Plan that advocates for cross-border trusted data spaces and closer alignment of national AI strategies. With respect to foreign judgments’ influence in China, Chinese courts do not recognise foreign judgments as binding precedent, but certain decisions have been cited in academic commentary and may provide reference to judicial reasoning in novel AI-related cases.

While foreign entities enjoy statutory IP enforcement rights under the National Treatment principle, they face significant administrative hurdles in respect of data and market access. Foreign investment in generative AI services in China may be subject to the applicable foreign investment restrictions. The transfer of data collected within China to a foreign jurisdiction is tightly regulated under PRC law, and a security assessment may be required.

The Interim Measures define “generative AI services” as “services offered to the public that generate text, images, audio, video or other content by GAI technologies”. Beyond this, however, there are no statutory definitions of core terms such as “AI system”, “foundation models”, or “general-purpose AI models”, “computer-generated works”, “AI-assisted inventions”, “training data”, “weights”, “prompts” or “outputs” in China. In practice, Chinese courts usually apply the traditional standard for copyright protection – namely, the requirement of human intellectual input – to determine whether any AI-generated output qualifies as a “work”. In the patent context, the China National Intellectual Property Administration (CNIPA) revised its Examination Guidelines (effective 1 January 2026) clarifying that “an AI system cannot be listed as an inventor”. This expressly confirms that inventorship remains reserved for natural persons.

China’s current laws and regulations do not define or distinguish “autonomous AI systems”, “agentic AI” or “AI agents” from other AI systems for IP purposes. However, the Interactive AI Measures specifically regulate the continuous emotional interactive services offered to the public that simulate the personality traits, thinking patterns and communication styles of a natural person through AI technologies, including – without limitation – emotional care, companionship and support provided in the forms of text, images, audio, video and the like.

Underlying the above is a structural tension in the Chinese approach. IP statutes, AI-specific regulation and regulatory guidance do not share a single set of definitions; indeed, there is currently no cross-referential definitional framework among them. Rather, the IP system stretches traditional concepts to cover AI outputs, while the AI regulatory framework defines a narrow class of AI services in technical terms – definitions that are not, in turn, imported into IP law. The result is a two-track regime in which the scope of protectable subject matter and the scope of regulated conduct are ascertained under separate and unaligned rules.

China does not maintain a dedicated body for AI-specific legislation, enforcement or adjudication. Instead, AI-related issues fall within the jurisdiction of the general court system, IP offices, regulatory authorities and other competent bodies that already administer IP-related matters:

• courts – in addition to the ordinary courts, there are specialised Internet Courts (Beijing, Hangzhou, Guangzhou) possessing jurisdiction over internet-related copyright disputes, which have emerged as the primary testing grounds for AI copyright cases, notably handling landmark lawsuits concerning ownership of AI-generated content;
• CNIPA handles patent and trade mark examination;
• the National Copyright Administration (NCAC) handles copyright registration; and
• the Cyberspace Administration of China (CAC) serves as the primary regulator for algorithms and data security, with authority to enforce regulations requiring AI providers to train models on legitimate data and label AI-generated content.

Notably, there is no specialist AI-specific IP procedure that replaces or stands apart from the general IP framework. AI-related IP disputes and administrative matters are therefore processed through the same procedural channels as conventional IP cases, with the existing institutions applying and adapting general rules to the distinct questions raised by AI.

In the PRC, different elements of an AI system may be protected under different legal regimes, including copyright, patent, trade secret and anti-unfair competition laws.

Software, Documentation and APIs

Source code, object code and software documentation may be protected as computer software under PRC laws. Such protection generally covers the expression of the software, rather than underlying ideas, algorithms or functions. Software code, technical documentation and certain system information may also be protected as trade secrets. The backend source code implementing an API may similarly be protected as software. In addition, PRC courts have recognised that unauthorised access to or exploitation of APIs and related data may constitute unfair competition under the Anti-Unfair Competition Law.

Model Architecture and Parameters

The source code implementing model architecture may be protected by copyright, while technical solutions relating to model design, optimisation, training or parameter adjustment may be eligible for patent protection. Model architecture and parameters themselves are generally not considered copyrightable works. However, where they are maintained as confidential technical information, they may be protected as trade secrets.

Prompts and Agent Configurations

Most prompts are unlikely to qualify for copyright protection because they generally consist of instructions or concepts rather than original expressions. In practice, commercially valuable prompts and agent configurations are more commonly protected through confidentiality measures as trade secrets.

Training Datasets and Data Assets

Training datasets may qualify as copyright-protected compilations where the selection or arrangement involves sufficient originality. Even where copyright protection is unavailable, datasets involving substantial investment in collection, processing or organisation may be protected as competitive interests under the Anti-Unfair Competition Law. Datasets containing confidential technical information may also qualify as trade secrets.

Under the PRC copyright framework and as discussed previously, copyright protection is limited to the original expression embodied in the software and does not extend to underlying ideas, algorithms, methods, technical principles or functionality.

Software and Model-Related Materials

Source code, object code and related software documentation may be protected as computer software; however, the underlying model architecture, algorithms and operational logic generally fall outside the scope of copyright protection. Depending on the circumstances, such elements may receive protection as competitive interests under the Anti-Unfair Competition Law.

Chinese courts have held that prompts consisting mainly of ideas and instructions lacked sufficient personalised expression to qualify as copyrightable works. However, prompt libraries may potentially qualify as compilations where the selection or arrangement involves sufficient originality.

In the PRC, AI-related inventions are generally subject to the same patentability requirements as other inventions. CNIPA has incorporated AI-specific considerations into the Patent Examination Guidelines. The latest amendments, effective from 1 January 2026, further clarify the examination standards for inventions involving AI, algorithms and data processing.

Technical Solution and Technical Effect

Under the PRC Patent Law, an invention must constitute a new technical solution relating to a product, process or improvement thereof. An abstract algorithm or AI model itself is generally not patentable unless it is incorporated into a technical solution that solves a technical problem and produces a technical effect.

The Patent Examination Guidelines provide that, for inventions involving algorithmic features, algorithmic features and technical features should be considered as an integrated whole. Where such features functionally support each other and produce an interactive effect, their contribution to the technical solution should be taken into account in assessing inventive step.

Disclosure and Claim Drafting

AI-related inventions must satisfy the general enablement requirements under the PRC Patent Law. The specification should disclose sufficient technical information to enable a person skilled in the art to implement the claimed solution. For AI inventions, this may include relevant aspects such as model architecture, training methods, input and output data relationships and parameters.

In practice, claims should define a technical solution incorporating AI features and identify the technical problem addressed, technical means adopted and technical effect achieved.

Under the PRC Anti-Unfair Competition Law, AI models, weights, datasets, prompts, system instructions, evaluation data and deployment know-how may be protected as trade secrets where they constitute technical or business information that is not publicly known, has commercial value and is subject to reasonable confidentiality measures.

Reasonable Confidentiality Measures

To maintain trade secret protection, AI developers are expected to implement confidentiality measures proportionate to the value and sensitivity of the relevant information. Common measures include:

• identifying and classifying confidential information;
• restricting access on a need-to-know basis;
• implementing encryption and access controls;
• maintaining access logs; and
• adopting technical measures to prevent unauthorised extraction or reverse engineering.

Companies typically also use confidentiality agreements and implement personnel management procedures, which may require more sophisticated technical and organisational controls due to the volume and complexity of the information involved in AI systems.

Transparency and Regulatory Disclosure

In the PRC, providers of generative AI services may be required to submit certain information, including algorithm-related materials, to regulatory authorities for filing or compliance purposes. Such disclosures are generally made to competent authorities rather than the public and, therefore, would not ordinarily result in the loss of trade secret protection.

The PRC does not currently provide a standalone database right. Instead, AI-related datasets, databases and other data assets may be protected through a combination of copyright, trade secret, contractual protection and unfair competition law, depending on their specific characteristics.

Copyright Protection

A dataset or database may qualify for copyright protection as a compilation work where the selection, arrangement or organisation of data reflects sufficient originality. However, copyright does not protect the underlying data itself, nor does investment in data collection alone establish copyright protection.

Trade Secret and Confidentiality Protection

Where datasets, annotations, labels, embeddings or related AI development materials are not publicly available and provide commercial value, they may be protected as trade secrets if the statutory requirements are satisfied.

For AI developers, commercially valuable information may include not only the datasets themselves but also data collection methods, data structures and other know-how developed during model training.

Protection Under the Anti-Unfair Competition Law

Data-related competitive interests may be protected under the Anti-Unfair Competition Law. A recent Supreme People’s Court case recognised that data collection created through a substantial investment and business operation could constitute a protected competitive interest, even though the data collection itself lacked sufficient originality to qualify for copyright protection.

Legal Certainty

Under the PRC copyright framework, the use of copyrighted works through activities such as copying, scraping, ingestion, analysis, tokenisation or transformation remains subject to existing copyright rules. It is legally required that training data be obtained from lawful sources and that the use of such data does not infringe IP rights or personal information rights.

Legal Uncertainty

The key unresolved issue is whether and to what extent AI developers may use copyrighted works without authorisation for AI training purposes. Article 24 of the Copyright Law of the PRC provides an exhaustive list of statutory exceptions, but AI model training is not expressly included. Therefore, there is currently no clear statutory basis for a general copyright exception permitting the use of copyrighted works for AI training.

Judicial Developments

PRC courts have begun to address IP issues arising from AI-related activities. In the “Hangzhou Ultraman case”, the court held that training may be deemed fair use if it is not aimed at reproducing the original work’s expressive elements, does not affect the normal exploitation of the work, and does not unreasonably prejudice the rights-holder’s legitimate interests. The “Shanghai Medusa case” further clarified the distinction between training and output. The Shanghai Intellectual Property Court held that merely inputting copyrighted images into a low-rank adaptation (LoRA) model for training does not necessarily infringe the reproduction right; rather, reproduction liability may arise where the trained model subsequently generates outputs that reproduce the protected expression. Overall, the Chinese judiciary’s approach to AI training copyright is one of case-by-case balancing, and a comprehensive framework governing the use of copyrighted works during AI training has not yet been established.

Existing Exceptions and Limitations

Under the PRC copyright framework, there is currently no specific statutory exception for text and data mining or AI model training. Accordingly, commercial AI developers generally cannot rely on a specific copyright exception when using copyrighted works for model training or deployment.

Judicial Developments and Remaining Uncertainty

In the “Ultraman LoRA case” (2024), the Hangzhou Internet Court indicated that the use of copyrighted works at the training stage may potentially qualify as fair use where the works are used as analytical samples to extract patterns or features, rather than to reproduce their protected expression, and where such use does not interfere with the normal exploitation of the works or unreasonably prejudice the rights-holder’s interests. Although the case did not establish a general AI training exception, it suggests that PRC courts may have some interpretative room to recognise certain analytical or non-expressive training uses as permissible under the existing copyright framework.

The availability of existing copyright exceptions may also depend on the purpose and context of use. Article 24 contains exceptions for personal study and research, and for certain limited uses in classroom teaching or scientific research. However, their statutory conditions make them difficult to apply to commercial AI training, and their application to AI training more generally remains unsettled.

AI developers in the PRC generally rely on existing copyright licensing arrangements to obtain rights to use training content.

Although copyright collective management organisations exist in certain sectors, such as music and literary works, no established collective licensing framework currently enables AI developers to obtain broad licences for training content at scale. Similarly, there are currently no AI-specific mechanisms addressing orphan works or out-of-commerce works for training purposes. Therefore, developers seeking to use such materials for AI training generally need to rely on existing copyright ownership rules and obtain licences where possible.

In practice, AI training content in the PRC is increasingly obtained through negotiated commercial arrangements with content owners, data providers or specialised dataset suppliers. Publicly available information regarding remuneration structures remains limited, and commercial terms are generally negotiated on a case-by-case basis, taking into account factors such as the uniqueness and quality of the dataset, volume of content, permitted uses, exclusivity, duration and compliance obligations.

At present, the PRC does not have a statutory opt-out mechanism equivalent to the commercial TDM opt-out regime under the EU Digital Single Market Directive. In practice, rights reservations and technical notices may serve as evidence of rights-holders’ intentions, contractual restrictions or factors relevant to assessing an AI developer’s knowledge and conduct. However, there is currently no established legal rule confirming that AI developers are required to detect and comply with such notices, or that failure to do so would automatically constitute copyright infringement.

Looking forward, the legal significance of rights reservations may increase if more standardised mechanisms are developed, including machine-readable declarations, content credentials, licensing registries or other systems enabling AI developers to identify restricted content. Where rights reservations are sufficiently accessible and technically identifiable, failure to respect such reservations may become a relevant factor in assessing copyright liability or the reasonableness of an AI developer’s conduct.

Provenance and Transparency Obligations

Under the Interim Measures for the Management of Generative Artificial Intelligence Services, providers of such services are required to complete relevant filing procedures and provide information regarding algorithm mechanisms, including relevant information on training data sources and data processing methods. Service providers may also be required to co-operate with regulatory inspections, which may involve providing information regarding training data sources, data labelling rules and algorithmic mechanisms.

However, Chinese law does not currently require public disclosure of complete training datasets. There is also no AI-specific statutory framework requiring developers to maintain records of all model development activities or agentic system operations. In practice, AI developers increasingly adopt internal data governance measures to manage IP and regulatory risks.

Cross-Border Training and Infringement Assessment

As a member of the Berne Convention, the PRC provides national treatment protection to copyrighted works originating from other member states. Accordingly, the use of foreign copyrighted works in AI development is generally assessed under the applicable PRC copyright framework where Chinese law applies.

In cross-border scenarios, relevant factors may include:

• the location where copyrighted works are accessed or copied;
• the location of relevant servers and development activities;
• the location of the AI service provider; and
• the place where infringement effects occur.

The assessment is likely to depend on the specific facts of each case.

Under the current Chinese legal framework, establishing direct copyright infringement by a model developer or provider in connection with training, fine-tuning, evaluation, deployment or operation requires the claimant to prove four elements under the fault-liability principle set out in the Civil Code – ie, that:

• the claimant holds copyright or an exclusive licence (or other enforcement standing) in the work at issue;
• the defendant carried out an act regulated under the Copyright Law, such as reproduction, adaptation or communication to the public via information networks;
• the act was unauthorised and does not fall within a statutory limitation on rights (such as fair use); and
• there is a causal link between the act and the resulting harm.

Chinese judicial practice currently establishes infringement primarily by working backwards from the generated output, rather than by directly evaluating the training process itself.

The Hangzhou Ultraman decisions suggest a more permissive approach to data input and model training but stricter scrutiny of generated outputs and their use. Pending cases involving Trik AI and MiniMax may clarify whether using copyrighted works for training can itself constitute infringement. In addition, the Supreme People’s Court is reportedly drafting judicial policy guidance on AI-related disputes – a development expected to bring greater doctrinal clarity to this currently unsettled area.

Chinese courts have not yet squarely determined whether the weights, parameters, embeddings or other internal artefacts of a general-purpose AI model, considered in themselves, constitute a “copy” or an adaptation of a particular copyrighted work. The key question is whether they embody identifiable, recoverable protected expression. Relevant evidence may include whether simple prompts can reliably reproduce the work, whether outputs are repeatable, and sufficiency of technical evidence on the model, training data or logs.

Chinese copyright decisions have not yet adopted “memorisation”, “regurgitation” or “de minimis copying” as distinct AI-specific legal tests. These concepts may nonetheless be evidentially relevant.

Where copying must be inferred from circumstantial evidence, Chinese courts commonly use the “access plus substantial similarity” approach. That is not, however, an exhaustive or inflexible test for every AI claim. Direct evidence of ingestion, storage, training, caching or retention may independently establish reproduction; and the exact enquiry will vary with the right invoked, the alleged act and the available technical evidence.

Under PRC law, a model provider may incur liability for IP infringement by users or downstream autonomous systems under general tort and network-service rules, subject to the safe harbour protection granted to network service providers, rather than under an AI-agent-specific regime. The principal test is fault: whether the provider has directly participated in infringement, or has knowingly or negligently facilitated it, and whether it has met a duty of care proportionate to its technical capability.

The general safe harbour principle in the PRC provides that, upon receiving a qualified notice, a network service provider must promptly forward it to the alleged infringer and take necessary measures (eg, deletion, blocking); failure to do so triggers joint liability for the expanded harm. Immediate joint liability may be imposed if the provider “knows or should know” of infringement yet fails to act. For generative-AI providers, the Interim Measures on Generative AI Services impose additional upstream duties – lawful training data, respect for IP rights, and a duty to act as a “content producer”. Courts assess liability based on the provider’s degree of control, direct profit, prominence of the IP, and whether it took technically feasible preventative measures beyond mere post-notice deletion.

When confidential materials are used to train, fine-tune, prompt, store in memory, retrieve (including via agents) or access through tool calls, PRC law provides remedies mainly through the Anti-Unfair Competition Law and confidentiality obligations under the Civil Code’s contract provisions. A trade secret claim requires proving three statutory elements: secrecy, commercial value, and reasonable confidentiality measures taken by the rights-holder.

The “fingertip recognition case” handed down by the Supreme People’s Court was the first to confirm that deep-learning training code and annotated databases can qualify as protectable trade secrets. It held that, even where an AI product’s front-end functionality is publicly accessible, secrecy is unaffected so long as strict confidentiality measures (eg, GitLab access controls, cloud encryption) protect the back-end training logic.

Where a model “reproduces, infers or discloses” confidential information during evaluation or deployment, this can trigger concurrent breach-of-contract liability (where contractual confidentiality clauses apply) and tort liability under the Anti-Unfair Competition Law.

No published Chinese court decision has yet directly addressed novel scenarios such as prompt injection causing confidential leakage, or an AI agent autonomously disclosing confidential material after retrieval. These issues are expected to be clarified in future AI-related trade secret disputes.

Substantive Defences

Fair use

The Copyright Law lists closed statutory exceptions that do not expressly address AI training, but the “Hangzhou Ultraman case” has shown a tendency to extend fair use to the training stage where training does not reproduce protectable expression and does not unreasonably harm the rights-holder’s normal exploitation or legitimate interests.

Lack of substantial similarity

Since Chinese courts generally apply “substantial similarity between the generated output and the original work” as the core test for infringement, a defendant may argue that the generated content does not substantially resemble the original work.

Safe harbour

AI model training technology is itself neutral, and a platform operator that has fulfilled a reasonable duty of care and promptly takes necessary measures appropriate to the service type may avoid contributory liability.

Independent creation

Under PRC copyright law, independent creation is not a standalone defence but operates within the “access + substantial similarity” test. Where the defendant proves no access to the plaintiff’s work and the output results from genuine independent effort, infringement cannot be established. In the AI context, the user must demonstrate meaningful human creative input.

Procedural Defences

Procedural defences available include the statute of limitations under the Civil Code (the general three-year limitation period) and jurisdictional objections – for instance, extraterritorial jurisdiction defences where cross-border cloud-based training activity is involved.

The authors have not seen PRC law enumerate defences such as implied licence, exhaustion, lack of substantial taking, public interest and abuse of rights or competition-law arguments generally as specifically codified defences. That said, the authors do not rule out the possibility that PRC courts may adopt or infer defences in nature similar to these.

Substantial Similarity in AI-Generated Outputs

In the PRC, courts generally assess copyright infringement by AI-generated outputs using the traditional “access + substantial similarity” approach. This approach has been reflected in Chinese AI-related copyright discussions and cases, including text-to-image disputes.

Prompts and Reference Materials

The same principle may apply where protected material is reproduced through prompts or reference images. If a user deliberately instructs an AI system to imitate another work’s style while also recreating its core plot, composition or other original expressive elements, this may be treated as an active instruction to reproduce the protected work. If the resulting output is substantially similar to that work, it may infringe copyright. In serious circumstances, criminal liability may also arise.

Style Imitation Without Literal Copying

Style imitation without literal copying requires a more careful distinction. Chinese copyright theory and practice generally follow the traditional idea–expression dichotomy: abstract ideas, artistic styles, genres, themes and techniques are not protected by copyright as such. However, where a supposed “style” is embodied in a particular work through identifiable and original expressive elements, an AI output that reproduces those elements may still infringe copyright.

Under PRC law, a user, deployer or customer can still face infringement liability for generating, using, publishing or commercialising AI output, or by configuring or deploying an agentic system that then carries out the infringing act – even if they genuinely did not know the output was infringing. That said, “not knowing” is not legally irrelevant. It can significantly affect the amount of damages, whether punitive damages apply, whether liability is shared with others, and whether the conduct rises to the level of a criminal offence.

Similarity remains the fact-based test for infringement itself – it applies regardless of anyone’s mental state, meaning even accidental or single-use generation of substantially similar output can be infringing.

Intent and knowledge mainly affect the degree of fault: deliberately feeding in a protected work suggests intentional infringement; a generic prompt producing an unexpected match may suggest no fault. This shapes damages, whether punitive damages apply, joint liability with developers, and criminal exposure.

Prompt design is the key evidence for proving intent and causation.

Human review, content filters and complaint procedures are important but not universally mandatory evidence of reasonable care; courts weigh the cost of such measures against their practical benefit.

Reliance on provider terms (disclaimers, ownership clauses) mainly governs the contractual relationship between user and platform and generally cannot block a third-party rights-holder’s claim.

Scale of use does not affect whether infringement occurred, but larger-scale, repeated or commercial use increases damages and raises the risk of “serious circumstance” findings that may trigger punitive or criminal liability.

In the PRC, AI outputs are judged by the same standards as human-made content: if a mark or brand identity is used to signal product origin and risks confusion, it is trade mark infringement or passing off under the Anti-Unfair Competition Law.

Unauthorised use of a celebrity’s likeness, voice or name in AI-generated endorsements violates Civil Code personality rights and can trigger liability under the Advertising Law and the Consumer Protection Law for violations against endorsement regulation.

AI involvement is never treated as a defence – liability turns on confusion, deception and unauthorised use, not on how the content was made.

When an AI system generates product designs, technical instructions, software code, manufacturing parameters, chemical or biological candidates, or processes, that output can fall within the scope of a patent or a design – and doing so can create infringement risk.

Direct Infringement

At the direct-infringement level, if the technical solution or product design that an AI system generates is identical to, or the technical equivalent of, the features claimed in someone else’s valid patent, whoever puts that solution into practice can be held liable for patent infringement, unless a statutory defence under the PRC Patent Law applies.

In practice, an AI system simply producing a design, a piece of code or a set of parameters is usually just the starting point of the risk, not the infringement itself. It is when that output is actually put into production, used or commercially exploited that direct infringement is most likely to be triggered.

Indirect Infringement

At the indirect-infringement level, the PRC judicial interpretations and the joint-liability rules under the Civil Code provide the main pathway for addressing AI-related scenarios.

The most distinctive challenge with AI-related patent infringement is that it tends to involve multiple parties and multiple steps. Bringing a technical solution to life typically spans the AI developer, the platform operator and the end user – and no single one of these parties, on their own, carries out every technical feature recited in the patent claim.

Chinese legal scholarship and judicial practice are currently exploring ways to work around this problem through causation-based analysis. The core emerging standard appears to examine whether a party’s conduct played an “irreplaceable and substantial role” in bringing about the infringement.

The PRC has not yet enacted a comprehensive law dealing specifically with liability for AI systems or autonomous AI agents. If an AI agent independently scrapes or retrieves third-party content, uses software tools or APIs, generates code or content, uploads material, or makes product or design choices, PRC law will generally apply existing rules on tort, contract, IP and online-platform liability.

Since an AI system is not a legal entity, it cannot itself be sued or held legally liable. Any liability must therefore be attributed to the people or companies behind it – for example, the AI provider, platform operator, system deployer, application developer, or end user. The fact that the AI acted “autonomously” does not automatically remove human responsibility.

Human Contribution and Evidentiary Requirements

Under the PRC Copyright Law, in assessing whether AI-generated output is eligible for copyright protection, courts examine whether the output constitutes an intellectual achievement bearing originality traceable to identifiable human creative activity. Recent decisions suggest an increasingly granular approach placing growing weight on evidence documenting whether and how the human contribution shapes the AI-generated output. Courts focus on whether the claimant can substantiate the underlying process and show that modifications, selections, edits made to prompts, parameters and intermediate outputs reflect personalised choices that shaped the final expression.

Cases denying protection show that, where a claimant cannot produce contemporaneous generation or editing records, the claimant cannot cure that evidentiary deficiency merely by using similar prompts during the proceedings to generate a new output resembling the disputed one. Such a reproduction would not be deemed to demonstrate that the claimant’s own contribution shaped and controlled the original output. Basic descriptive prompts, single-shot generation, and unmodified acceptance of AI output likewise do not evidence sufficient human contribution to the resulting expression.

Agentic Systems

As for agentic systems, China currently has no dedicated rules or representative published case on the copyrightability of AI agent-generated content. Because agentic systems can perform tasks with a greater degree of autonomy, the human contribution to the resulting output may be less identifiable, and it may accordingly be more difficult to draw a clear line between protectable human expression and unprotectable AI-generated material within such integrated outputs.

As discussed in 6.1 Human Authorship and Copyright Protection, sufficient human contribution is the decisive factor that courts apply in determining whether AI-generated output qualifies for copyright protection. Where an output lacks a human author or reflects only minimal human contribution, courts will find it unprotectable as a work even if it has been registered as a copyrighted work. PRC law does not currently establish a separate right for such unprotected output.

The PRC Copyright Law has not established a standalone copyright regime for AI outputs; ownership shares, the qualification of adaptations or transformations as derivative works, and the permissions required to use existing material are all governed by the existing general copyright framework rather than any AI-specific rules.

Where an AI output involves multiple contributors, ownership shares are accordingly determined by each contributor’s substantive contribution to the output, are applied under the general co-authorship rules, and are in practice usually fixed by contract rather than left to this statutory default. Where the output constitutes an adaptation or transformation of an existing work, it may itself qualify for copyright protection as a derivative work provided it reflects substantive human contribution. That said, exercising that protection remains subject to, and cannot infringe, the rights in the underlying work, which in turn means that any adaptation of, or reliance on, third-party material still requires the relevant rights-holder’s prior authorisation, save in the very limited fair use circumstances recognised under PRC law.

Under PRC law and based on the authors’ anonymous consultation with the competent authority, applicants are not subject to a standalone requirement to disclose AI involvement for the purpose of registering or enforcing copyright or patent rights. Moreover, AI usage need not be stated in the relevant application documents, nor are applicants separately required to submit any declaration regarding the use of AI, although separate rules require certain AI-generated content to carry labels identifying it as AI-generated when generated or disseminated online.

PRC law does not currently prescribe any AI-specific consequence for failing to make such disclosure in relevant applications. It remains uncertain, however, whether non-disclosure in particular cases could constitute a breach of other generally applicable requirements such as the good faith principle. In certain circumstances, such non-disclosure may trigger tort or even more severe liabilities (for example, a claimant alleges copyright ownership in an AI-generated piece of output that is actually not copyrightable and makes profit from licensees who believe in the claimant’s alleged authorship).

Under the PRC Patent Law, inventors must be natural persons, and an AI system therefore cannot be named as an inventor or co-inventor. For AI-assisted inventions, a natural person may be named as an inventor only if they have made substantive creative contributions to the technical features of the invention. However, PRC law currently does not provide specific criteria for assessing such contributions in the AI context, and the determination remains fact-specific. The use of AI tools does not create a separate entitlement regime; ownership of inventions involving employees, contractors, researchers or collaborators continues to be determined by the existing PRC Patent Law framework and applicable contractual arrangements.

Inventive Step and Enablement

In December 2024, CNIPA issued the Guidelines for Patent Applications for AI-related Inventions (for Trial Implementation), categorising AI-related applications into four types:

• AI algorithms or models themselves;
• functional or domain applications based on AI algorithms/models;
• AI-assisted inventions; and
• inventions involving AI-generated content.

The Guidelines provide specific guidance on inventive step and enablement for AI-related inventions within the existing patent examination framework. For inventive step, AI-related algorithmic features that are functionally integrated with technical features should be considered as a whole. For enablement, applicants must provide sufficient disclosure to support implementation of the claimed invention; however, the appropriate level of disclosure for AI-related inventions, including model structures, training data and optimisation methods, remains subject to practical challenges.

Skilled Person and Prior Art

CNIPA has not issued official guidance on whether AI availability should enhance the capabilities attributed to the notional skilled person. In the current examination practice, however, AI does not replace the skilled-person assessment and may only serve as an auxiliary reference rather than an independent basis for inventive-step findings. Whether AI-generated outputs constitute prior art also remains unsettled. Although no official rules exist, commentators advocate caution given the risk that large-scale AI-generated materials may distort prior art searches if indiscriminately treated as citable prior art.

China has not adopted separate rules for AI-generated designs, graphical user interfaces (GUIs), icons, avatars, virtual goods or trade dress. AI involvement therefore neither precludes nor independently establishes protection, which remains governed by the existing design patent, copyright, trade mark and unfair competition regimes.

Product configurations and qualifying GUIs may obtain design patent protection if they are suitable for industrial application, new, and clearly distinguishable from prior designs or combinations of prior design features. A GUI must also be embodied in a product and involve human–computer interaction; game interfaces are ineligible for patent application. A partial design consisting solely of a pattern, or a combination of pattern and colour, on a product’s surface is likewise ineligible.

Original visual expression in icons, avatars and virtual goods may be protected by copyright where it reflects identifiable human intellectual input. Source-identifying visual elements may also be protected under trade mark law. Trade dress is not a standalone right, but packaging, decoration, application names and icons with a certain degree of influence may be protected against confusing use under unfair competition law. Ownership and enforcement follow the ordinary rules applicable to each regime; AI itself cannot be a designer, author or rights-owner.

Registrability

Under the current PRC Trade Mark Law, a trade mark must be distinctive, must not fall within any prohibited or non-registrable category, and must not conflict with prior trade marks or other prior rights. Neither the creator’s identity, the method of creation, nor human authorship is an independent registration requirement. Therefore, in theory, AI-generated names, logos, slogans and sounds may be registered if the resulting signs satisfy the ordinary requirements. The 2026 revision to the same law, which will take effect on 1 January 2027, expressly adds motion marks to the list of registrable signs.

Ownership

As regards ownership, only natural persons, legal persons and other organisations may apply for registration. An AI system falls within none of these categories and cannot apply for or hold a registration in its own name. Ownership of an AI-assisted brand asset therefore vests in the applicant recorded as registrant once approved, regardless of AI’s role in generating it.

PRC Copyright law grants human authors moral rights, including the rights of publication, attribution, alteration and integrity, and grants performers analogous personal rights to be identified and to protect their performance image against distortion. For an AI-assisted work, attribution follows the author recognised under the Copyright Law. Although statutory labelling rules require disclosure of AI generation or synthesis, these requirements concern transparency rather than authorship.

The same logic governs alteration and integrity. Where an output reproduces and materially alters or distorts the protected expression of an existing work without authorisation, the author may invoke the right of alteration and integrity, alongside any economic-rights claim. Similarly, if a synthetic AI performance is created by substantially altering an existing recorded performance, the performer may invoke the right to be identified and the right against distortion.

As for false attribution, where content is falsely credited to a real person who did not create or perform it, this may infringe that person’s right of name and right of attribution, and may also infringe their reputation right if the association damages their social standing. Where content is simply presented as human-created without naming anyone, the concern shifts to public confusion, which in commercial contexts may also raise issues under the AI-labelling rules.

Under PRC law, an individual’s name, portrait, voice, persona or digital replica is protected against AI generation or use primarily through the right of name, right of likeness, right of voice, which are a cluster of rights similar to the right of publicity in the common law system, as well as the right of reputation and personal information protection rules.

Protection of Names, Images and Voices

Courts have consistently held that unauthorised use of a person’s name, portrait or voice to create AI-generated content, such as a synthetic voice or an AI character, infringes the corresponding right of likeness and right of voice. Where AI combines several personality elements to appropriate an individual’s overall persona in a way no single right fully addresses, general protection for personal dignity could also apply.

Protection of Personal Information

Generating such content typically also involves processing the underlying biometric data, which triggers a further, independent layer of protection: sensitive personal information such as facial images or voiceprints requires the individual’s separate, specific consent before use for AI purposes, and unauthorised processing, including for face-swapping, has been held to infringe personal information rights.

Protection of Performance

Performers’ rights under the Copyright Law include the right of attribution and the right of protection against distortion, and they would apply where AI distorts a performer’s existing performance. Where AI instead generates a performance the performer did not actually give, these rights under the Copyright Law would generally not extend to it, with personality rights providing the relevant protection instead.

As mentioned in 1.4 Courts, IP Offices and Regulators, settlement on AI-related IP disputes is subject to the same forum as for general IP cases in China.

For the purpose of obtaining evidence concerning training data, model development, prompts, logs, outputs, source code, model weights, filters or evaluation results, including tool-call histories, browsing records, API logs, agent instructions, memory stores and orchestration logs, rights-holders may draw on the same procedural tools available in general IP litigation.

• Pre-litigation evidence preservation: rights-holders may apply to the court for preservation of evidence before filing a lawsuit, where evidence is likely to be destroyed or to become difficult to obtain later.
• Court-ordered evidence production during litigation: courts may order a party to produce evidence in its control that the opposing party cannot otherwise obtain. If the controlling party fails to comply without justifiable grounds, the court may draw adverse inferences against it.
• Investigative orders from courts: a party that cannot obtain evidence due to objective reasons can apply to the court to investigate and collect it. The court can then issue orders to third parties – including cloud service providers, API platform operators, or AI service platforms – to produce relevant records. 

Several mechanisms are available to protect trade secrets and confidential technical information from disclosure in the course of litigation.

• Private hearings: cases involving trade secrets may be heard in private upon application by a party, restricting public access to sensitive technical material.
• Limited cross-examination: where evidence involves trade secrets, cross-examination shall not be conducted openly after the evidence is submitted, limiting exposure of confidential information during examination of the evidence.

Injunctive relief available in AI and IP disputes is derived from the traditional civil litigation framework in China.

Interim injunctive relief (so-called “behaviour preservation” under the Civil Procedure Law, supplemented by the 2018 SPC Provisions on Behaviour Preservation in IP Cases) may be granted before final judgment. It functions to preserve the status quo and prevent irreparable harm while proceedings are pending. Courts apply a four-part test in deciding whether to grant such relief, balancing:

• likelihood of infringement;
• irreparable harm;
• balance of preservation interest and damage to the respondent; and
• public interest.

The foregoing is subject to the applicant posting security to cover the respondent’s potential losses. Interim injunctive relief is typically conditioned on the applicant posting security to cover the respondent’s potential losses.

Final injunctive relief is issued only after a full trial on the merits establishes liability, and operates primarily to compel the infringing party to cease the infringing conduct. In the AI context, such an order may encompass halting training on infringing data, discontinuing deployment or distribution of the model or service, and disabling infringing outputs. In addition, at the request of the infringed party, the court may direct the infringing party to publish corrective notices or a public apology in order to eliminate the continuing impact of the infringement.

Remedies for AI and IP claims mirror the framework of traditional IP litigation, encompassing both monetary and non-monetary measures. In assessing monetary damages, courts will first determine the actual losses suffered by the infringed party or the profits derived by the infringing party from the infringement. Where neither the actual losses nor the infringing party’s profits can be quantified with sufficient certainty, courts will have recourse to a reasonable royalty or licence fee as the basis of assessment. If even a reasonable royalty or licence fee proves difficult to quantify, courts may, on a case-by-case basis, award statutory damages of up to RMB5 million. Crucially, wilful infringement in serious circumstances exposes the infringing party to punitive damages of up to five times the base monetary damages as determined in the foregoing. Beyond monetary awards, Chinese courts place considerable weight on non-monetary remedies, principally the final injunction as noted in 8.3 Interim and Final Injunctive Relief.

China strictly adheres to the principle of territoriality for IP rights. Tort claims regarding IP infringement are governed by the law of the place where the infringement occurred (lex loci delicti) or the law of the forum (lex fori). The enforcement of foreign judgments or orders in China requires the existence of a treaty or reciprocal arrangement with the rendering jurisdiction. Even under established frameworks – such as the landmark Mainland-Hong Kong Reciprocal Enforcement Arrangement – interim relief and final injunctions for IP cases are explicitly excluded from cross-border enforcement, subject to a narrow exception in respect of trade secrets.

In the PRC, AI-related content licences are generally structured through negotiated contractual arrangements, with parties tailoring the scope of permitted use according to the specific AI application and commercial objectives. The key provisions outlined below are for reference only and are not exhaustive.

Scope of Licence and Permitted AI Uses

A licence should clearly identify the licensed materials, which may include copyrighted works, databases, datasets, annotations, metadata and other content resources, and should distinguish between different permitted uses, including model training, fine-tuning, evaluation, retrieval-augmented generation (RAG), agentic retrieval, tool-based access and user-facing display.

Data Sources

Licences and related data supply agreements should address the source and ownership of the licensed materials, the scope of rights obtained by the licensor, and the licensor’s ability to grant the relevant AI-related permissions.

Term, Exclusivity and Remuneration

Key commercial terms also include duration, territory, exclusivity and remuneration. Parties may also need to consider whether remuneration should be structured as a fixed licence fee, usage-based fee, subscription fee or other commercial arrangement, depending on the nature of the AI application and the scale of content use.

Deletion, Opt-Outs and Downstream Outputs

More and more AI content licences address deletion requests, withdrawal mechanisms, and the treatment of content already incorporated into AI systems. Agreements should also allocate rights and responsibilities relating to downstream outputs, including ownership, permitted commercial use, infringement risks and indemnification obligations.

Several key legislative and regulatory initiatives are pending or expected in China, including a comprehensive national Artificial Intelligence Law, judicial guidelines by the Supreme People’s Court on AI-generated content and copyright, and revisions to the Implementing Regulations of the Copyright Law. The People’s Courts Implementation Plan for Judicial Protection of Intellectual Property Rights (2026–2030), issued by the Supreme People’s Court on 17 April 2026, specifically directs that AI-related cases be handled to promote “beneficial, safe and equitable development”. Courts are instructed to accurately determine the legal attributes of AI-generated content – evaluating whether output reflects the author’s originality of selection and expression based on the specific instructions entered and the process of selecting and modifying content. The Implementation Plan calls for balancing development promotion with regulatory oversight, prudently hearing new-type cases involving large-model training-corpus use and AI-content infringement, and exploring judicial rules on the ownership of rights in AI-generated works while precisely defining the responsibilities of AI developers, operators and users.

Taken together, these pending reforms may be broadly favourable to the rights-holders and would make AI-related IP enforcement more predictable and effective. AI developers would benefit from a more predictable compliance baseline and clearer safe-harbour and duty-of-care boundaries, while in exchange assuming heavier clearance responsibilities for training image, voice and textual data, more concrete liability-exposure calculations, and precisely defined accountability across the developer-operator-user chain. For AI users, the reforms would more clearly delineate rights and responsibilities, strengthening content-safety and labelling protections and clarifying ownership of AI-generated works and derivative-use rights.

To date, China has maintained a close policy dialogue with the World Intellectual Property Organization (WIPO) on improving international IP standards with respect to AI through the organisation’s Beijing office. For instance, China presented submissions on WIPO’s draft Issues Paper on IP Policy and AI, the organisation’s first major attempt to codify IP policy with respect to AI. Similarly, China has submitted multiple interventions to the organisation’s Conversations on IP and Artificial Intelligence. Going forward, China intends to take on an even more active role in this regulatory area. The deputy head of CNIPA has stated that China, during the period of the 15th Five-Year Plan (2026–2030), will “deeply engage in international co-operation on AI-related IPR [intellectual property rights], [and] actively participate in the formulation of relevant international rules, technical standards, and governance frameworks”.

The recent creation of WAICO, a forum for international co-operation on the development of AI, likely represents a further opportunity for China to assume a greater leadership role in the co-ordination of AI and IP laws. A central principle of WAICO is enhanced international alignment on “AI development strategies, governance rules and technical standards, so as to form a consensus-based global governance framework”.

Among others, the copyrightability of AI-generated work is the main point of divergence between China and other major jurisdictions. As early as 2023, for instance, Chinese courts began to recognise copyright in AI-generated images where reflective of an author’s personalised expression. While the US Copyright Office originally took a different approach, rejecting copyright registration for works generated by AI without substantial alteration by a natural person, it now recognises copyright where AI has been used in an assistive capacity to enhance human expression. In this way, the US position has hewed closer to China’s, albeit while maintaining different analytical bases.