Major Asian jurisdictions are responding to the intellectual property challenges posed by generative artificial intelligence (particularly copyright challenges) through specialised rules and judicial decisions. Different jurisdictions are developing distinct approaches to training data use, fair use, ownership of generated content and platform liability. This article first examines legislative and judicial developments in South Korea, Japan and Singapore, then considers the emerging body of Chinese case law, before turning to copyright issues associated with the development and operation of AI models that existing rules have yet to address adequately.

Legislative and Judicial Developments in Representative Asian Jurisdictions

South Korea, Japan and Singapore are addressing copyright issues arising from generative AI through copyright limitations and exceptions, administrative guidance and judicial decisions. However, the scope and intensity of protection vary.

South Korea principally relies on the traditional fair use doctrine. Guidelines issued in 2026 state that storing works during model training constitutes reproduction in principle. Whether such use qualifies as fair use requires consideration of the purpose and character of the use, the nature of the work, the amount used and the effect on the relevant market. Factors favouring fair use include the transformative nature of the training, its contribution to the public interest and the adoption of technical safeguards designed to prevent the model from generating outputs that reproduce the original work. AI-generated outputs are generally not protected as works; only those elements reflecting identifiable human creative contributions may be eligible for protection. The National Institute of Korean Language Corpus case further demonstrates that a public-interest purpose does not, by itself, eliminate the need to obtain authorisation and that conventional publishing authorisations do not necessarily encompass the right to use works for AI training.

Article 30-4 of Japan’s Copyright Act establishes a relatively permissive exception for the use of works for “non-enjoyment purposes.” In principle, general data analysis may therefore be conducted without authorisation. The exception does not apply, however, where fine-tuning or overfitting is conducted with the purpose of reproducing a particular work or where the use unreasonably prejudices the work’s existing or potential licensing market. Infringement by AI-generated outputs continues to be assessed by reference to similarity and dependence. The copyrightability of generated content depends on whether a human user has creative intent and exercises creative control over the final expression through specific instructions, iterative revision, selection and arrangement or substantive modification.

Singapore’s regime is centred on the “computational data analysis exception.” It permits the use of works where the user has lawful access, the copies are used solely for computational data analysis purposes and such copies are not made available or distributed beyond what is permitted. The exception applies only to the training stage and does not automatically shield infringing outputs from liability. Courts will consider whether an output reproduces a substantial part of the original work and whether the user, developer or deployer controlled the relevant acts, knew or ought to have known of the risk and adopted reasonable preventive measures. The unauthorised use of corporate project data for AI training may also give rise to trade secret liability.

Overall, Japan and Singapore provide comparatively clear statutory frameworks for AI training through specific copyright exceptions. At the same time, South Korea relies more heavily on case-by-case analysis under its fair use framework. All three jurisdictions increasingly focus on the lawful provenance of training data, market substitution and the role of technical safeguards in mitigating infringement risks, while balancing technological innovation against the interests of right holders.

Emerging Rules in Chinese Judicial Practice

In the absence of comprehensive AI-specific legislation, China currently addresses intellectual property disputes arising from generative AI products and services primarily through the Copyright Law, the Civil Code, the Anti-Unfair Competition Law and the Interim Measures for the Management of Generative Artificial Intelligence Services. Through individual cases, the Beijing, Hangzhou and Guangzhou Internet Courts, together with specialised intellectual property courts, are gradually shaping preliminary judicial approaches concerning the use of training data, ownership of generated content and platform liability.

Copyrightability of AI-generated content

First, on the copyrightability of AI-generated content, Chinese judicial practice is gradually developing an approach centred on “human creative contribution.” In the Beijing Internet Court’s AI-generated image case, the court held that an image could qualify as an artistic work where the user formulated prompts, adjusted parameters, generated and selected outputs through repeated iterations and thereby made personalised arrangements in the image’s expression. The central question was not whether AI had participated in the process, but whether the final expression could be attributed to the user’s original intellectual contribution. However, this does not mean that merely inputting prompts is sufficient to establish copyright protection. In the Suzhou “Butterfly Chair” case, the court held that prompts merely combining simple concepts, without differentiated expression and control over the expression, were insufficient to establish originality. Similarly, in the Shanghai “prompt” case, the court found that prompts consisting of conventional industry expressions and lacking individualised characteristics did not themselves constitute copyrightable works. Taken together, these decisions indicate an emerging judicial focus on whether the user has made an individualised, recognisable and substantive contribution to the final expression.

The Beijing Internet Court’s virtual digital human case further demonstrates that copyright protection remains focused on human creative expression. A virtual digital human not directly derived from a real person may qualify as an artistic work if it reflects the production team’s distinctive aesthetic choices in lines, colours, form and overall visual design. A virtual digital human driven by a real person or modelled on the image or characteristics of a real person, may also implicate portrait rights, voice rights and performers’ rights. The AI-generated voice and AI companion cases adopt identifiability as their central criterion. Where a synthetic voice, portrait or virtual persona enables the public to associate it with a particular natural person, its unauthorised commercial use may infringe personality rights. Although these rules do not fall within copyright law in the narrow sense, they directly affect the lawful use of AI training materials and generated content. In a typical case published by the Supreme People’s Court, the Court likewise confirmed that an AI-generated voice identifiable as belonging to a particular person by reference to its timbre, intonation and manner of pronunciation falls within the protection of that person’s voice rights.

The use of training data and the application of fair use remain among the most contentious issues

China’s Copyright Law does not contain a specific text and data mining exception and AI training is not expressly included among the fair use circumstances enumerated in Article 24. Accordingly, no uniform answer has emerged as to whether the unauthorised reproduction of works for the commercial training of AI models may qualify as fair use. Driven in part by the need to accommodate industrial development, judicial practice has begun to distinguish among the data-input, model-training and content-output stages. At the training stage, the analysis focuses on whether the purpose is to extract patterns rather than reproduce protected expression and whether the use interferes with the normal exploitation of the work or unreasonably prejudices the legitimate interests of the right holder. Because the output stage directly produces and disseminates content, courts place greater emphasis on substantial similarity and market substitution.

The “Medusa LoRA Model” case illustrates this layered approach. The court found that LoRA parameters generally constitute a statistical compression of features extracted from works rather than a digital encoding of the original images. Completion of the training process therefore does not necessarily mean that a copy of the original work has been created. According to the court, infringement of the reproduction right may arise only where the model is capable of reproducing the original work and actually generates an output substantially similar to it.

The case distinguishes “memorisation” in the technical sense from “reproduction” within the meaning of copyright law. It does not, however, conclusively answer whether the unauthorised input of protected works for training purposes constitutes fair use. The Hangzhou “Ultraman” case also adopted a stage-specific approach, but its primary focus was platform liability. After a user uploaded protected materials to train a model, the platform failed to take necessary measures against the infringing model and its outputs; therefore, it was held contributorily liable for infringement of the right of communication through information networks. In our view, it would be premature to characterise this decision as generally recognising AI training as fair use.

Liability

Finally, platform liability is shifting from formal reliance on “technological neutrality” towards a substantive review of the platform’s actual role. Courts consider whether the platform participated in model training or content generation, induced infringement through algorithmic design, exercised control over outputs or derived commercial benefits from the relevant service. A platform that merely stores user content may remain eligible for protection under the notice-and-takedown regime. By contrast, where its algorithms actively organise, encourage or directly generate infringing content, the platform may incur direct or contributory liability. The AI companion case, the Ultraman case and the AI-generated “product recommendation post” case collectively demonstrate that courts may also consider product functions, business models and effects on the broader content ecosystem, supplementing copyright rules with personality rights and unfair competition law where appropriate.

Generative AI disputes

Overall, Chinese courts have developed a preliminary adjudicatory framework for generative AI disputes centred on human creative contribution, the distinction between training and output stages, substantial similarity and the degree of platform participation. These individual decisions, however, do not yet constitute uniform rules of general application. Important questions remain unresolved, including whether the use of training materials may qualify as fair use, whether model training creates a “reproduction” within the meaning of copyright law and how the precise boundaries of platform duties of care should be defined. The Supreme People’s Court is currently drafting an opinion on the proper adjudication of AI-related disputes and has proposed further exploration of rules governing ownership of AI-generated content and the respective liabilities of the parties involved. Relevant adjudicatory standards may therefore become more consistent over time. This development also gives rise to a more fundamental question: even where a specific output does not reach the threshold of substantial similarity, if the model’s weights or parameters can reliably imitate a particular work or an individual creator’s style, should copyright law intervene? If so, how should such intervention be structured?

Beyond Training Data: Emerging Copyright Liability at the Model Level

The cases discussed above largely address the legality of training inputs, the copyrightability of AI-generated content and liability for substantially similar outputs. A more difficult question concerns the AI model itself. Where no particular output is substantially similar to a protected work, but the model’s architecture, weights or parameters have been adapted to imitate a particular creator’s works or style, should – and at what point – copyright law intervene?

The question goes to the boundary between protected expression and the public domain. Under the idea–expression dichotomy, copyright protects expression rather than ideas, methods or styles. Article 9(2) of the TRIPS Agreement confines copyright protection to expressions, while Section 102(b) of the US Copyright Act excludes ideas, procedures, processes, systems and methods of operation. Although China’s Copyright Law does not expressly codify this dichotomy, the principle follows from the definition of a “work” under Article 3 and Article 6 of the Regulations on Computer Software Protection and established judicial practice.

Generative AI complicates this distinction by shifting the object of inquiry from visible expression in a particular output to the parameter space that shapes a model’s behaviour. The following discussion considers this emerging issue at three levels: the model, the model’s behaviour and evidentiary proof.

The model layer: code, weights and the meaning of “copy”

Copyright traditionally protects expression fixed in a perceptible or reproducible form. A trained neural network, by contrast, consists of vast arrays of numerical weights. Whether those weights can embody protected expression (or constitute a reproduction of the works used to train the model) remains unresolved.

Fine-tuning and low-rank adaptation or LoRA, make the question particularly concrete. A relatively small parameter file trained on a limited collection of images may cause a base model to generate outputs that consistently resemble the works or style of a particular creator. The parameter file does not display the source images and may not contain conventional digital copies of them. Nevertheless, it can exert a stable and reproducible influence over the model’s output distribution.

Two competing characterisations are possible. If model parameters are treated as a digital mapping or transformed fixation of protected expression, reproducing or distributing the parameter file could itself raise copyright concerns, even before a substantially similar output is generated. If the parameters merely encode unprotectable concepts, statistical correlations or stylistic characteristics, model-level liability becomes much harder to establish unless protected expression is reproduced in a particular output.

Courts have begun to confront this issue without resolving it. In Andersen v Stability AI Ltd., the US District Court for the Northern District of California allowed two direct-infringement theories to survive the pleading stage. Under the “model theory,” the AI model itself was alleged to constitute an infringing copy because protected works remained embodied within it in some form. Under the “distribution theory,” the plaintiffs alleged that distributing the model distributed the works incorporated into it. The court held only that these theories had been plausibly pleaded and expressly left their factual and legal validity for later determination.

The litigation in Doe 1 v GitHub, Inc. illustrates the opposite difficulty. The district court dismissed claims concerning the removal of copyright management information on the basis that the challenged output was not identical to the original code. Whether an “identicality” requirement is justified is still being contested on appeal. These cases illustrate that model parameters occupy a legal grey area: they may be too closely linked to protected expression to be considered mere ideas, but are often too abstract and technically altered to fit traditional definitions of copies.

Chinese courts have so far approached the issue principally through infringement and platform liability rather than by defining the copyright status of model parameters. As discussed in Part II, the Hangzhou “Ultraman LoRA” case imposed liability because user-trained LoRA models could be used to generate images substantially similar to the protected character and the platform failed to take necessary measures. The “Medusa LoRA Model” case more directly considered the technical nature of LoRA parameters, but stopped short of establishing a general rule that a parameter file either is or is not a reproduction. Whether model weights themselves can constitute infringing copies therefore remains open under Chinese law.

The behaviour layer: autonomous generation and the attribution of liability

Generative models are probabilistic rather than deterministic. Even a facially neutral prompt may produce content that reproduces protected expression. This phenomenon, commonly described as “memorisation” or “regurgitation,” raises a question that cannot always be answered by examining the user’s conduct alone: should an infringing output be treated as an isolated malfunction or as evidence that the model was designed, trained or deployed in a manner that predictably facilitates infringement?

The complaints in The New York Times Co. v Microsoft Corporation presented numerous examples in which OpenAI systems allegedly reproduced or closely paraphrased protected news content. The court subsequently allowed significant direct and contributory infringement claims to proceed while dismissing certain “abridgement,” DMCA and unfair-competition claims. The decision did not determine ultimate liability, but it confirmed that repeated examples of model reproduction may support plausible allegations concerning the provider’s role in user infringement.

US doctrine offers two established reference points. Under Sony Corporation v Universal City Studios, a technology capable of substantial non-infringing uses is not ordinarily subject to secondary liability merely because it can also be used to infringe. Under MGM Studios, Inc. v Grokster, Ltd., however, liability may arise where the provider promotes infringement through affirmative conduct or clear expressions of intent. Generative AI sits uneasily between these principles. A general-purpose model may have extensive lawful uses while still being trained, configured or marketed in ways that facilitate imitating particular works or creators.

Chinese law approaches this issue through fault, control and the duty of care. Article 1197 of the Civil Code imposes joint liability where a network service provider knows or should know that a user is infringing and fails to take necessary measures. The Supreme People’s Court’s judicial interpretation of the right to communicate through information networks further elaborates this standard.

In generative AI cases, relevant factors may include:

• the provider’s business model;
• the prominence of the protected work;
• the obviousness and frequency of infringing outputs;
• the provider’s technical capacity to intervene;
• the cost and effectiveness of alternative safeguards; and
• the broader effect of liability on innovation.

A structural tension is likely to matter increasingly. Providers may invoke technological neutrality in litigation while assuring enterprise customers, through contracts, marketing materials or product documentation, that outputs will not infringe third-party rights. They may also advertise the model’s capacity to imitate identified artists or styles. In Andersen v Stability AI Ltd. allegations concerning lists of artists associated with Midjourney became relevant to claims that the service had been marketed through the identities and reputations of particular creators.

A provider that gives extensive non-infringement assurances arguably acknowledges some capacity to control model behaviour; the greater that control, the narrower the practical scope of a neutrality defence may become. Conversely, if the system is genuinely uncontrollable, the reliability of such assurances becomes questionable. Future litigation is therefore likely to examine not only the model’s outputs, but also the relationship between its technical design, commercial representations and contractual allocation of risk.

The evidence layer: from discovery to technical deconstruction

A theory of algorithmic infringement, however well constructed, must withstand evidentiary scrutiny. Demonstrating that model parameters embody protected expression may require access to the system’s internal components, including model weights, training logs and data-lineage records. These are precisely the materials that defendants are likely to protect as core trade secrets.

Two procedural mechanisms are likely to shape this area. The first is the use of protective orders for trade secrets. In US litigation, tiered protective orders under Rule 26(c) of the Federal Rules of Civil Procedure, including “attorneys’ eyes only” designations, have become the standard compromise. The consolidated proceedings against OpenAI further demonstrate that discovery may itself create collateral risks: an order requiring the preservation of user conversation logs transformed a routine evidentiary measure into a large-scale privacy dispute. Chinese law provides a comparable toolkit. The Supreme People’s Court’s 2020 Provisions on Evidence in Civil Intellectual Property Litigation establish confidentiality measures and empower courts to order the production of evidence controlled by the opposing party, with adverse consequences for unjustified non-compliance. A distinct difficulty nevertheless remains: unlike a document, a model may be intelligible only when executed. Whether a protective order remains workable when examination requires running the model and who should supervise that process, remain unresolved.

The second mechanism is expert evidence. Courts will require assistance from specialists in AI interpretability and computational linguistics to explain, in legally cognisable terms, the functions of individual neurons, network layers and low-rank matrices. Chinese civil procedure accommodates this need through judicial appraisal and the participation of persons with specialised knowledge, while US practice addresses it through the Daubert gatekeeping framework. Expert disputes will ultimately concern what “similarity” means in parameter space: whether a style is present because:

• it was included in the training data;
• it can be elicited through targeted prompts; or
• it measurably shapes the model’s output distribution.

Looking ahead, courts may develop a test for algorithmic expression, much as patent law constructed (and subsequently constrained) the category of business-method patents through State Street, Bilski and Alice. Software copyright offers a possible template in the abstraction–filtration–comparison test articulated in Computer Associates v Altai. An analogous framework for models might filter out unprotectable ideas at the architecture level and compare the residual expression at the learned-mapping level. No court has yet adopted such a test. The doctrinal history of business-method patents, which proceeded through expansion, retrenchment and eventual stabilisation, nevertheless suggests one possible trajectory of development, if not its destination.

Reserving a doctrinal interface for algorithmic originality

Whether liability should attach to the outputs of a generative system or to the system itself ultimately depends on where copyright’s boundary is drawn in the generative era. Current legal developments provide answers mainly at the margins: courts calibrate duties of care, adapt evidentiary mechanisms and test competing theories claim by claim, while leaving the core question open. For now, evidentiary procedures, contractual risk allocation and dynamically calibrated duties of care may provide the most workable interfaces between copyright liability and algorithmic systems. The questions continue to outnumber the answers, which accurately reflects the present state of the law.