1. The systematic construction of the EU’s TDM exception clauses

The current EU legal framework governing TDM and copyright is characterised by a progression from a principled foundation towards increasingly detailed and specific obligations. The EU’s system of copyright exceptions and limitations, founded upon Article 5 of the InfoSoc Directive, constitutes a closed and exhaustive statutory list. This reflects the civil law tradition of authors’ rights, which emphasizes legal certainty. The DSMD introduced a significant update to this system by establishing a dual-track exception regime: research organisations benefit from a mandatory exception, while commercial applications are subject to a rights reservation mechanism.Footnote ¹⁴ It should be noted that the DSMD does not provide a clear definition of TDM,Footnote ¹⁵ leading some to question whether AI training activities fall within the scope of the statutory concept of TDM.Footnote ¹⁶ Subsequently, the enacted AI Act further established a risk-based regulatory framework, imposing transparency and copyright management requirements on general-purpose AI models.Footnote ¹⁷ Furthermore, the AI Act further explicitly stipulates that all providers of general-purpose AI models must comply with EU copyright law. This allows AI model training activities that meet the exception criteria to be exempt from copyright infringement liability, thereby providing essential legal certainty for AI research and development while respecting copyright protections. Moreover, the AI Act extends the regulatory scope, imposing clear requirements particularly on the training and use of general-purpose AI models. It shifts copyright compliance responsibilities earlier to the model development phase,Footnote ¹⁸ reflecting a regulatory shift from ex post infringement determinations to ex ante compliance guidance. Building on this foundation, the EU recently introduced the Code of Practice for General-Purpose AI (GPAI CoP), which provides granular interpretation of the statutory duties under the AI Act. Although this Code falls within the category of “soft law,” it complements the aforementioned “hard law” instruments, collectively forming a comprehensive regulatory chain covering data usage, model development and output dissemination.

Of particular note are the relevant provisions of the AI Act, which took effect in August 2025 and are poised to substantially impact copyright disputes related to generative AI. The Act strengthens transparency obligations; while this does not alter the legal nature of potential infringements, it makes them considerably easier to identify. Model providers are now required to publish detailed summaries of their training data and implement opt-out mechanisms for rightsholders. The heightened compliance burden significantly increases operational costs for businesses.Footnote ¹⁹ All providers of general-purpose AI models must conduct rigorous due diligence on the copyright status of training data and establish mechanisms to respect rightsholders’ reservations. For non-profit organisations like LAIONFootnote ²⁰ or SMEs such as HowardsHome,Footnote ²¹ the cost of building and maintaining a compliant system can be prohibitively high. Consequently, even though the statutory exception for TDM remains applicable in theory, the practical transparency and data provenance requirements may stifle non-profit, distributed, community-sourced data projects.

The EU’s legal provisions concerning TDM are grounded in the InfoSoc Directive. Through the interplay of the DSMD, the AI Act and the GPAI Cop, a progressively stringent and detailed regulatory framework has been formed. The InfoSoc Directive provides the foundational norms for copyright exceptions and limitations; however, it did not explicitly include TDM within the scope of such exceptions, effectively leaving room for the judicial application of copyright exceptions to TDM. Building upon this foundation, the DSMD excluded TDM conducted for non-scientific research purposes from the scope of copyright exceptions unless rightsholders had not expressly reserved their rights. This shift has created a significant dilemma for AI training for commercial purposes, introducing substantial legal risk. The AI Act does not, in essence, alter the legal position of TDM as established by the DSMD; it reiterates that providers of AI models must comply with the “opt-out” reservation of rights expressed in Article 4(3) of the Directive. Furthermore, the AI Act requires AI model providers to draft and implement a policy aimed at respecting Union copyright law, notably through the use of state-of-the-art technical means. This subjects commercial TDM to more rigorous legal oversight. For TDM, the GPAI Cop can effectively be viewed as a detailed elaboration of this technical requirement stipulated by the AI Act. Observing the trajectory of this regulatory development reveals the EU’s intention to mitigate the risks TDM poses to the copyright system through increasingly stricter regulation.

Under the new framework, the central controversy is shifting. It is moving away from substantive debates over whether an act constitutes infringement, and towards uncertainties regarding procedural compliance. For instance, while the AI Act mandates that companies publish a “sufficiently detailed” summary of training data, what constitutes “sufficiently detailed” remains undefined. The GPAI CoP offers best-practice guidance rather than hard standards. Although nominally voluntary, the AI Act stipulates that providers not adhering to the GPAI CoP must demonstrate compliance through alternative means approved by the European Commission.Footnote ²² In practice, this makes adherence to the Code the lowest-risk option for companies, potentially transforming voluntary norms into de facto mandatory standards. Ultimately, the legal risk for businesses is not eliminated but transmuted: from the substantive risk of infringement to a procedural and technical compliance risk, centred on the adequacy of processes and disclosures, judged against ambiguous and evolving standards. This shift in risk profile confirms the academic view that relying on copyright exceptions to justify AI training faces fundamental challenges. The EU’s regulatory approach effectively steers the industry toward a compliance model centered on licensing.Footnote ²³

Despite the EU’s intention to establish a comprehensive regulatory regime for TDM, its practical implementation has diverged from the legislative intent.Footnote ²⁴ The EU continues to face multiple challenges in reconciling TDM with copyright protection.Footnote ²⁵ On one hand, ambiguities within the legal provisions, insufficient coordination among Member States and a lack of uniform standards for machine-readable formats have hindered the effective enforcement of rights reservations. On the other hand, high compliance costs pose a barrier for small and medium-sized enterprises, whilst regulatory fragmentation increases the operational burden for multinational corporations. At the Member State level, divergent stances have emerged due to varying industrial interests, with a perceptible tension between the preferences of digital economy powerhouses and culturally prominent nations, further complicating policy alignment. Additionally, external pressures, including influence from the United States and industry lobbying, may lead to the delay or dilution of certain rules, exacerbating the risk of internal market fragmentation. Analysing the impact on different stakeholders, whilst rightsholders have gained new theoretical avenues for remuneration, they still encounter practical difficulties in exercising their rights. Research institutions, although permitted to use data under exception clauses, may suffer indirect losses due to constraints within the commercial ecosystem. Collectively, these factors test the EU’s capacity to balance technological innovation with copyright protection. In the short term, regulatory uncertainty is likely to persist. In the long term, a failure to achieve effective coordination could adversely affect the EU’s position in the global AI competition.

2. Case studies on TDM exception disputes in the EU

Despite maintaining a unified framework at the legislative level, the practical implementation of EU rules exhibits significant divergence among Member States, reflecting underlying coordination deficits and legal uncertainty at the Union level.Footnote ²⁶ Although some studies suggest that TDM complies with the “three-step test,”Footnote ²⁷ its practical application remains contested in academic discourse. German courts have demonstrated a strict, dichotomous approach, heavily emphasising the purpose of use when applying exceptions. In LAION v Robert Kneschke, the Hamburg Regional Court ruled that the creation of a dataset for non-profit research purposes fell within the TDM exception for scientific research under German copyright law.Footnote ²⁸ Conversely, in the case brought by GEMA against OpenAI, the Munich court, citing the model’s commercial nature, declined to apply the more lenient research exception. Instead, it invoked the general exception provision and found infringement due to the rightsholder’s prior reservation of rights.Footnote ²⁹

When determining infringement, German courts consider not only the nature of the use but, more critically, its impact on the potential market for the original work. In GEMA v OpenAI, ChatGPT’s output of song lyrics was deemed a direct reproduction of the work’s core content, conflicting with the normal exploitation of the work and thus constituting substantial use. This judicial inclination is rooted in Germany’s long-standing author’s rights tradition, demonstrating a value orientation that tends to favour the protection of rightsholders’ interests when balancing technological development and copyright.

An examination of rulings from German courts reveals that, based on the legal position of TDM established by the DSMD, the focus of litigation concerning TDM has shifted from the traditional question of whether an action constitutes a copyright exception to whether it serves a scientific research purpose. While this shift circumvents interpretive disputes over copyright exceptions found in traditional judicial decisions, it has led to a new institutional trap. Determining whether TDM activities in AI model training serve a scientific research purpose involves a significant degree of subjective judgement. This assessment of purpose effectively depends on whether the involved entity is non-profit and whether the AI in question is ultimately intended for commercial use.Footnote ³⁰ In reality, however, the commercial nature of an AI model’s training process is not strongly correlated with the commercial status of the entity undertaking it or the model’s subsequent deployment for commercial purposes.

By contrast, the Dutch court in the DPG Media v. HowardsHome case exhibited a different reasoning. The Amsterdam District Court did not automatically preclude the application of the TDM exception solely based on commercial character, but instead focused on the specific characteristics of the use. The court considered the provision of short text snippets accompanied by links to the original content an act of aggregation that neither substituted nor directly competed with the source material.Footnote ³¹ Furthermore, the court adopted a stricter view on technical specifics, ruling that a rights reservation must precisely correspond to the specific technical means employed to be legally effective. This reflects a Dutch judicial insistence on the principle of technological neutrality, rejecting broad-brush restrictions on automated processing.

The Dutch ruling displays a distinct pragmatism, employing quantitative standards to classify short excerpts as permissible and emphasising their referential, rather than substitutive, function. This seeks a balance between investor rights and the free flow of information. Unlike the German emphasis on protecting creators, the Dutch approach appears more inclined to preserve space for data flow and technological innovation. Collectively, these cases illustrate that when reconciling AI development with copyright protection, national courts within the EU often engage in nuanced, case-specific balancing of interests, resulting in a pluralistic landscape of judicial practice.

However, this landscape of fragmented jurisprudence is at a potential turning point. The Hungarian court has referred the case of Like Company v. Google Ireland to the CJEU, a move anticipated to foster a more harmonised legal interpretation at the EU level. The core legal issues revolve around whether text processing during model training constitutes a regulated act of reproduction, and if so, whether such training can fall within the TDM exception.Footnote ³² A ruling by the CJEU on these matters will transcend the individual case, providing an authoritative definition for pivotal concepts such as “reproduction” and the “TDM exception.”

The recently enacted provisions of the AI Act and the associated GPAI CoP, while not altering the legal standard for determining copyright infringement itself, are poised to exert considerable influence on judicial decisions through their transparency obligations and code of conduct mechanisms. The Act’s requirement for providers of general-purpose AI models to publish detailed summaries of their training data brings greater clarity to the factual matrix of disputes.Footnote ³³ Rightsholders may find it easier to demonstrate whether their works were included in a training dataset, while providers can leverage these public records to evidence their compliance measures. Should a provider have voluntarily adhered to the relevant code of practice, courts may consider this a factor demonstrating diligent effort, potentially favourably influencing liability assessments. Nevertheless, the broad definitions within certain provisions of the Act, coupled with potential variances in implementation across Member States, risk generating new divergences and further exacerbating legal uncertainty.

Tracing the evolution of relevant EU judicial practice reveals a discernible pivot from an initial encouragement of technological exploration towards a phase of intensified regulation. Early rulings adopted a more permissive stance towards TDM, aiming to nurture an emerging technological field. As the technology matured and its market impact expanded, more recent judgements have displayed a stronger inclination to reinforce copyright protection, signalling a corrective recalibration towards upholding creators’ rights. This trajectory underscores the enduring challenge for the EU in constructing its TDM framework: achieving an effective and sustainable balance between legal harmonisation, technological innovation and copyright protection.

3. Legislative motivations and policy directions behind the EU’s TDM rules

The EU’s policy trajectory in the realms of AI and copyright represents the outcome of protracted negotiations and compromise among multiple stakeholders.Footnote ³⁴ In fact, EU copyright law-making has not been driven by copyright-specific rationales, but rather by the objective of establishing an internal market.Footnote ³⁵ In the early days, online platforms extensively utilised copyrighted content under safe harbour provisions without providing fair remuneration to creators, sparking debates over an inequitable value distribution. Industry bodies, notably the Federation of European Publishers, actively lobbied for a legislative response.Footnote ³⁶ This prompted the EU to introduce copyright filtering obligations for online service providers within the DSMD, aiming to rebalance the interests through reinforced platform accountability.

Subsequently, the drafting process of the AI Act brought these tensions to a head. Initially, the legislation demonstrated the EU’s strategic intent to assert its digital sovereignty and uphold fundamental rights via high-standard regulation. By establishing what it envisioned as the world’s most rigorous regulatory framework, the EU sought to leverage its rule-making power as an asymmetric competitive advantage, thereby steering the global development of AI to align with its values and market rules. However, the legislative process encountered substantial resistance from industry. US tech firms, led by OpenAI, openly questioned the feasibility of certain provisions,Footnote ³⁷ at one point threatening to exit the European market,Footnote ³⁸ while engaging in intensive lobbying efforts. Simultaneously, divisions emerged within the EU itself, with Member States such as France and Germany expressing concerns that excessive regulation could stifle the growth of domestic AI enterprises like Mistral AI.Footnote ³⁹

Under this dual pressure, legislators ultimately made significant concessions.Footnote ⁴⁰ The final version of the Act abandoned the initial concept of automatically categorising all General-Purpose AI systems as high-risk, opting instead for a new regulatory tier that imposes stricter obligations solely on the most powerful models presenting systemic risks. The evolving stance of OpenAI exemplifies a broader pragmatic shift in corporate attitude. Its approach transitioned from highly publicised threats of withdrawal in 2023 to stating, following an unsuccessful legal challenge in 2025, that it was considering an appeal.Footnote ⁴¹ This indicates a growing focus on navigating operational space within the compliance framework, rather than outright confrontation. This evolution signifies that the relationship between regulation and innovation has entered a new stage, where the emphasis pivots from adversarial debates on principle prior to legislation to the practicalities of compliance adaptation and dispute settlement thereafter.

Transparency requirements, championed by rights holders, have also been incorporated into the legislation, mandating that generative AI systems disclose copyright information related to their training data. This directly addresses concerns raised by the creative industries. The final version of the Act emerges as a hybrid outcome of competing interests, where various demands have been partially accommodated. While the technology sector faces compliance pressures, it has also averted more extreme regulatory measures. However, it is worth noting that stringent regulations centred on building data barriers could inadvertently increase compliance costs for domestic companies, reducing the diversity and quality of AI modelsFootnote ⁴² and ultimately constraining the long-term competitiveness of the EU’s AI industry.

Moving forward, businesses will need to navigate dual compliance requirements under both the DSMD and the AI Act. Such complexity may gradually lead the industry to develop best practices. As more companies adopt codes of conduct and national regulatory bodies within Member States become more established, the EU’s approach to AI governance could influence the formation of global standards through the Brussels Effect. Nevertheless, the divergence between the EU’s strict regulatory path and the US principle of fair use remains a significant obstacle to international regulatory harmonisation.

Recently, the European Commission has considered suspending certain provisions, including granting grace periods to companies and postponing transparency-related penalties.Footnote ⁴³ This reflects an ongoing trend of dynamic policy recalibration. Such moves respond not only to lobbying by US tech giants but also to internal calls from European industries seeking to avoid stifling innovation. These developments illustrate regulators’ efforts to strike a balance between external pressures and domestic demands, while also signalling a reassessment of the potential adverse effects of the initially stringent regulatory framework.

1. The “transformative use” standard in US law and its application

In addressing the copyright challenges posed by TDM, the judicial practice in the United States exhibits a distinct pragmatic inclination, deeply rooted in its fair use doctrine. Originating from the common law tradition of judicial precedent, this doctrine was codified in Section 107 of the Copyright Act of 1976.Footnote ⁴⁴ It is not a specific catalogue of rights but an open-ended principle of judicial equity. At its core, judges are required to conduct a comprehensive balancing of four flexible statutory factors on a case-by-case basis: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used and the effect of the use upon the potential market for or value of the work. This allows for certain unlicensed, yet socially beneficial uses, such as criticism, comment, news reporting, teaching, scholarship or research.

Compared to the EU’s exhaustive list of copyright exceptions, although both share the legislative objective of balancing the interests of rightsholders with public interests like freedom of expression, education and innovation – and both must comply with the internationally accepted three-step test – their legal logics and implementation paths are fundamentally different. The US fair use doctrine is a flexible mechanism combining rules and standards. It is inherently common-law in nature, serving as a safety valve grounded in utilitarian philosophy and the First Amendment. It is capable of proactively adapting to technological shifts – from videocassette recorders and search engines to AI training – embodying both forward-looking potential and inherent uncertainty. In contrast, the EU’s list of exceptions is a product of the civil law droit d’auteur tradition, manifested as a closed, exhaustive enumeration within statutory law. User conduct must strictly fit within a specific provision to qualify for an exemption. While offering greater legal certainty, it lacks the inherent flexibility to respond to novel developments.

This fundamental paradigmatic divide between common law judicial dynamism and civil law legislative rigidity is vividly illustrated in US judicial practice. Specifically, courts, by flexibly interpreting the transformative use standard within the fair use doctrine, have regarded AI training as a transformative use of the original works.Footnote ⁴⁵ The purpose is not merely to reproduce the content of the originals, but to learn their underlying patterns through algorithms to generate entirely new functionalities and value. This approach has, in effect, fostered a relatively permissive innovation environment for the AI industry. This judicial stance not only reflects the law’s adaptability to technological developments but also, on a deeper level, serves the strategic objective of enhancing the nation’s technological competitiveness.

Several landmark cases clearly illustrate the conditional support adopted by US courts regarding copyright issues in AI training data. In The New York Times Company v OpenAI Inc, the dispute extended beyond the training phase to the model’s output behaviour, indicating that the assessment of transformativeness requires a holistic consideration of both the technical process and the consequences of use.Footnote ⁴⁶ The ruling in Richard Kadrey v Meta Platforms Inc was more groundbreaking, as the court explicitly recognised that the training process itself constitutes highly transformative use, given its aim is to create tools capable of text comprehension and generation, rather than to disseminate the original expression.Footnote ⁴⁷ This judgment provided significant judicial support for technology companies using publicly available data for model training.

This judicial stance centered on transformative use becomes particularly salient when contrasted with the UK’s approach to similar disputes. In the cross-border parallel litigation of Getty Images v Stability AI, the legal paths of the two nations diverge markedly: on the US front, the core of the litigation revolves closely around the principle of transformative use; whereas in the UK, due to the strict technical definition of infringing copies under the law and jurisdictional hurdles, the focus of the lawsuit has been shifted towards trademark infringement, with the court having ruled that the AI model itself does not constitute a copyright-infringing copy.Footnote ⁴⁸ This contrast highlights the unique defensive space that the flexibility of the US fair use doctrine provides to AI companies, while the UK’s more traditional, text-based interpretation may impose more stringent evidentiary challenges on rights holders.

Of course, the US support for transformative use is not unconditional. The ruling in Kristen Bartz v Anthropic PBC introduced a crucial limitation, with the judge unequivocally stating that the legality of the data source is a prerequisite for claiming transformative use. Even if the training purpose is innovative, obtaining data through pirated channels inherently constitutes infringement.Footnote ⁴⁹ This position has prompted the industry to shift towards compliant data sources and spurred the emergence of a new data licensing market. Notably, companies of different sizes face disparate situations in this transition: well-resourced large firms can mitigate risks through bulk licensing agreements, whereas startups may struggle due to compliance barriers.

This judicial direction effectively constitutes an implicit industrial policy. Through a relatively permissive interpretation of transformativeness, US courts have created a regulatory environment for domestic AI enterprises that is more favourable than that faced by their EU counterparts. This competitive strategy is particularly evident in cases like the lawsuit by Westlaw’s parent company against the startup Ross Intelligence, where the court’s stringent ruling protected the interests of established content giants while inadvertently raising barriers to market entry.Footnote ⁵⁰ Furthermore, the U.S. Copyright Office has established the disclosure requirements for AI-generated content in copyright registration at the administrative level by issuing the Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence. At the same time, through the release of the Copyright and Artificial Intelligence series of reports, it has guided and stimulated deeper discussions on issues such as the legality of training data. The legality of training data sources is currently being addressed by Congress through legislative proposals like the Generative AI Copyright Disclosure Act and has not yet been fully or definitively clarified through administrative guidance.

The United States has currently developed a unique governance ecosystem: statutory legislation remains silent, the judiciary exerts influence through the accumulation of case law and de facto standards emerge commercially through market negotiations. This pluralistic dialogue mechanism maintains the legal system’s flexibility but also introduces significant uncertainty. Licensing negotiations between large technology firms and content providers are shaping new industry norms, with judicial rulings either endorsing or correcting these developments. This dynamic equilibrium continuously moulds the industrial landscape, fostering technological progress while potentially raising future antitrust concerns or triggering legislative intervention.

In summary, to consolidate and maintain its global leadership in AI, the US has adopted a strategy of flexibly interpreting the transformative use doctrine in its judicial practice. This approach establishes a resilient balance between copyright protection and technological innovation,Footnote ⁵¹ effectively creating a favourable innovation climate for domestic companies. This mechanism provides the necessary legal latitude for AI development while steering the industry towards compliance through market forces. Its fundamental rationale is to proactively serve the sustained development needs of strategic national industries through the adaptive interpretation of legal principles.

2. An examination of the status quo of the fair use system in China’s copyright law

China’s legal regulations and judicial practice in the field of TDM reflect profound policy considerations. As competition intensifies between Chinese AI enterprises and their international rivals, particularly leading US companies, legislators must carefully balance the relationship between copyright protection and technological innovation to avoid falling behind in the global race. China’s approach to applying its fair use rules to copyright issues arising from TDM demonstrates distinct pragmatism.Footnote ⁵² Presently, unlike the EU, China has not established a specific copyright exception for TDM. Legal determinations primarily rely on the general provisions for fair use within the Copyright Law,Footnote ⁵³ combined with a comprehensive assessment utilising legal instruments such as the Anti-Unfair Competition Law. This model reflects a pragmatic strategy that prioritises regulating the output end while adopting a relatively tolerant stance towards the input end. It aims to maintain market order and content security while preserving a degree of exploratory space for the development of the AI industry.

Several representative cases clearly illustrate the reasoning and standards applied by Chinese judicial practice in handling related controversies. China’s first copyright infringement case involving the training phase of a large-scale AI image generation model directly addresses the core issue of data acquisition for AI training. An illustrator accused the developer and operator of an AI painting software of using their publicly posted works to train the AI model without authorisation, pushing the judicial review focus forward to the input stage of training data. The core dispute lies in whether this act of reproduction for training purposes constitutes infringement.Footnote ⁵⁴ The plaintiff’s choice to sue directly over the training act itself, rather than acting only after infringing content is generated, indicates that content creators are beginning to actively assert their rights in the AI era. The Chinese court’s characterisation of the nature of this training activity in this case will directly influence the entire industry’s development model.

In the case concerning the Shuabao app crawling short videos from Tiktok, the Chinese court demonstrated its focus on maintaining fair competition order during the data processing and utilisation phases. From the perspective of corporate strategy, Tiktok’s decision to base its lawsuit on unfair competition rather than copyright infringement highlights the platform’s intent to protect the aggregate value of its data collection. The court ruled that the crawling constituted unfair competition, providing an alternative protective path for data-related interests through competition law.Footnote ⁵⁵ This approach sidesteps the debate over whether a data collection constitutes a copyrightable work while effectively curbing free-riding behaviour, reflecting a pragmatic judicial effort to uphold market fairness.

The world’s first ruling on the tort liability of a generative AI platform, issued by the Guangzhou Internet Court, exemplifies strict regulation of output-side conduct. In this case, the court held the AI platform operator liable for users generating images substantially similar to the “Ultraman” character using its service. The liability was based on the operator’s failure to implement necessary proactive screening measures and to fulfil a reasonable duty of care, constituting contributory infringement.Footnote ⁵⁶ Notably, the court supported the plaintiff’s claim to cease the specific infringing acts in question but rejected the request to delete all related training data from the platform.Footnote ⁵⁷ This indirectly indicates that the court distinguishes between data use in the training phase and infringing outcomes at the output stage, leaving interpretive space for training acts – not aimed at reproducing the expressive content of a work and not unreasonably prejudicing the legitimate interests of the right holder – to potentially be recognised as fair use.

The role of China’s Copyright Law in the field of TDM can also be observed in disputes that did not proceed to litigation. For example, China National Knowledge Infrastructure (CNKI) once sent a letter to Shanghai Mita Network Technology Co., Ltd., asserting that its copyright in compilations had been infringed, alleging that Mita AI’s search engine had crawled and displayed titles and abstracts of a large volume of academic literature without authorisation.Footnote ⁵⁸ The dispute was ultimately not taken to court but was resolved swiftly within 24 hours of Mita receiving the letter, through a public statement, removal of the relevant data and disconnection of links. The root cause of this outcome lies in the significant ambiguities within China’s Copyright Law regarding such novel data utilisation disputes, making it difficult for either party to ensure victory in court. For CNKI, the basis of its claim rests on treating the collection of numerous literature titles and abstracts as a protected compilation. However, this is challenging in judicial practice. Chinese courts have increasingly stringent requirements regarding the originality of selection and arrangement in databases. CNKI’s database emphasises comprehensiveness of coverage over creative arrangement, creating substantial uncertainty about whether it would be recognised as possessing sufficient originality. For Mita, the potential defence of fair use also carries high risks. Although its actions had a public-interest nature in facilitating academic research, under China’s Copyright Law, the key criteria for determining fair use include that the act “does not unreasonably prejudice the legitimate interests of the copyright owner” and is consistent with the “three-step test.” Faced with commercial, large-scale and systematic data crawling and substitutive display, these conditions are difficult to satisfy, and a court would likely find that such actions substantially harm the market value of CNKI’s database. Therefore, the ambiguity of the law creates a grey area of comparable risk for both parties. Litigation is not only time-consuming and costly but could also produce precedents unfavourable to one’s own side or the entire industry. Through a commercial warning and swift settlement, CNKI maintained its data barriers at a relatively low cost, while Mita averted existential threats and shifted to alternative data sources.

A review of these cases and regulatory practices reveals that China has adopted a multi-level, dynamically coordinated governance strategy in the field of TDM. At the top-level design level, regulations such as the Interim Measures for the Management of Generative Artificial Intelligence ServicesFootnote ⁵⁹ emphasise the responsibilities of service providers regarding content security, labelling obligations and copyright compliance, demonstrating a focus on regulating the output and application ends. The Guidance on Deepening the Implementation of the “AI Plus” Initiative further proposes improving data property rights and copyright regimes to align with AI development, and encourages the exploration of value-contribution-based data cost compensation and revenue sharing mechanisms.Footnote ⁶⁰ At the judicial level, courts exhibit caution, exploring boundaries on a case-by-case basis. The core consideration often centres on whether the use affects the normal exploitation of the work and unreasonably prejudices the legitimate interests of the right holder. At the enforcement level, initiatives such as the “Clear and Bright Governance of AI Technology Misuse” campaign led by the Cyberspace Administration of China exemplify a proactive regulatory approach. By reinforcing primary responsibility on enterprises and intervening at the source of training data management, authorities have urged companies to clean existing datasets and establish strict web-crawling standards, thereby translating principle-based requirements into concrete compliance guidance. Such administrative oversight also demonstrates flexibility and local adaptability: cyberspace authorities in regions such as Beijing, Shanghai and Zhejiang have adopted measures tailored to local industrial contexts – including multi-stakeholder governance, awareness-raising campaigns and technical blocking mechanisms – to jointly promote lawful and compliant TDM. This “output-focused, input-lenient” governance model, under the current circumstance of unclear copyright rules, objectively affords the AI industry a degree of flexibility and grey area regarding data acquisition and model training. It reflects a characteristically pragmatic wisdom of “crossing the river by feeling the stones.”

3. Comparison and analysis of the EU, US and Chinese models

In addressing the copyright challenges posed by TDM, the United States, the European Union and China have developed distinct governance pathways shaped by their respective policy priorities. These approaches not only reflect differences in regulatory design across the three major legal systems but also reveal deeper tensions among competing objectives such as technological innovation, cultural protection, and industrial security.

The EU’s legal and policy framework for TDM, centred on the DSMD and the AI Act, aims to balance technological innovation with copyright protection.Footnote ⁶¹ However, its underlying logic is marked by significant internal challenges. The “opt-out” mechanism in the DSMD, which styles itself as an exception yet is effectively regarded as a right that rightsholders must actively reserve to exclude material from training datasets,Footnote ⁶² and which allows rightsholders to exercise their rights at any time, results in a dynamic copyright status for high-value data. Moreover, the opt-out mechanism, reliant on non-standardised signals, exacerbates uncertainty and fails to ensure transparency or fair compensation.Footnote ⁶³ This uncertainty makes it difficult for businesses to ensure long-term compliance, thereby dampening incentives for innovationFootnote ⁶⁴ and potentially placing European AI developers at a competitive disadvantage globally.Footnote ⁶⁵ Concurrently, the AI Act’s requirement to publish summaries of training data to enhance transparency may conflict with the protection of corporate trade secrets. Technically, the accurate, real-time identification of copyright status within vast datasets is both costly and formidable.Footnote ⁶⁶ Supplementary voluntary codes of practice, due to their non-binding nature and reliance on widespread industry adoption, offer limited practical coordination. The layering of these rules also risks fragmenting the internal market, as divergent transposition by Member States increases legal complexity.

In the international context, the EU’s path contrasts sharply with the US and Chinese models. The US leverages its fair use doctrine, allowing courts to address TDM flexibly through case law. This approach, focused on promoting innovation and maintaining global technological leadership, fosters a relatively permissive environment for the AI industry.Footnote ⁶⁷ However, its reliance on ex-post judicial rulings creates legal uncertainty, and high litigation costs can stifle startups. The absence of a prior exemption mechanism and the practical impossibility of licensing vast datasets on a per-work basis leave businesses in a persistent state of legal ambiguity – a significant drawback the EU should avoid emulating. Nevertheless, the adaptive nature of the US fair use model is worthy of consideration, as it allows courts to adjust standards dynamically with technological progress, avoiding legislative rigidity and providing flexibility for innovation.

China adopts a pragmatic regulatory model, foregoing a specific TDM exception in favour of a comprehensive assessment utilising the general fair use clause in its Copyright Law alongside instruments like the Anti-Unfair Competition Law. This strategy focuses regulatory effort on the output end, allowing flexibility for copying activities at the input stage.Footnote ⁶⁸ While this provides the industry with operational leeway during its developmental phase, the ambiguity surrounding input activities means businesses may operate based on regulatory tolerance rather than legal certainty, potentially affecting long-term investment confidence. This inherent vagueness is a disadvantage the EU should not replicate, as it can lead to protracted instability. However, China’s emphasis on coordinating development with controllability offers a valuable insight for the EU. For instance, its approach of managing risks by reinforcing service provider obligations, rather than excessively restricting data input, can help foster technological iteration while safeguarding rights.

The future trajectory of the law remains uncertain. The EU may undertake de facto flexible adjustments through soft-law instruments like guidelines and promoted codes of practice, providing the industry with breathing room. The definitive interpretation of key legal concepts will likely fall to the CJEU, though this process is protracted and its outcomes unpredictable. Ultimately, EU policy will likely continue to oscillate between upholding rights protection standards and striving for AI leadership, suggesting that instability in its legal environment may persist in the short to medium term.Footnote ⁶⁹

1. Policy leadership: fostering a multi-stakeholder copyright governance ecosystem

To address the predicament faced by TDM in the application of copyright law, the EU must construct a multi-level, multi-institution collaborative governance framework. The core objective is to create a stable and affordable compliance environment for businesses, especially small and medium-sized enterprises (SMEs).Footnote ⁷⁰ Specifically, at the regulatory level, the European Commission should clarify exemptions based on turnover in the implementing provisions of the DSMD, allowing small businesses to temporarily avoid some of the strictest filtering obligations. To reduce transaction costs, the European Union Intellectual Property Office (EUIPO) could collaborate with collective management organisations in Member States to develop and promote standardised, simplified licensing agreement templates. Regarding financial support, programmes such as the Digital Europe Programme or dedicated AI innovation funds could be leveraged to provide direct subsidies for SMEs’ compliance expenditures. Additionally, exploring the inclusion of certain compliance costs within tax deduction schemes would substantially reduce the financial burden on enterprises from a fiscal perspective. While reducing costs for businesses, balancing the interests between rights holders and users is equally crucial. The European Commission could promote the extended collective licensing model through legislative initiatives to safeguard the actual income of rights holders. Moreover, utilising the EU Blockchain Services Infrastructure to develop a prototype micro-payment system based on distributed ledger technology could enable real-time rights verification and remuneration payment. Furthermore, the European Committee for Standardisation (CEN) should take the lead in advancing the standardisation of copyright information identification technologies.

Coordination at the governance level is also indispensable. Within the EU, binding implementation guidelines could be issued through regulatory structures established under the AI Act, such as the AI Office, to limit Member States’ discretion when transposing TDM exceptions into national law. The EUIPO could also establish and maintain a unified EU TDM rules database, clearly presenting regulatory differences among Member States and enhancing the transparency of legal application. At the international level, the EU should proactively utilise multilateral platforms such as the EU–US Trade and Technology Council (TTC) to explain the rationale behind its high-standard TDM rules and actively promote their external adoption. Simultaneously, the EU should advocate its balanced governance model in international forums like the World Intellectual Property Organization (WIPO) to foster global consensus. Additionally, active efforts should be made through European Committee for Standardization (CEN) and the International Organization for Standardization (ISO) to advance the global standardisation of copyright information identification technologies, thereby reducing cross-border data flow barriers arising from regulatory differences at the infrastructure level.

In summary, by comprehensively employing policy instruments such as differentiated regulation, economic incentives, governance coordination, and international dialogue, the EU has the potential to systematically resolve the copyright exception dilemma surrounding TDM, ultimately achieving a harmonious interaction between copyright protection and technological innovation.

2. Legal practice: recommendations for legislative refinement and judicial guidelines on TDM exceptions

To resolve the dilemma of applying copyright law to TDM, the construction of a theoretical framework must be grounded in transforming abstract legal principles into clear, actionable and specific rules. This systematic endeavour can be advanced along three closely interrelated dimensions:

First, efforts should focus on clarifying legal provisions to eliminate ambiguities. Specific measures could include the European Commission promoting standardised licensing request templates or endorsing industry-recognised copyright filtering tools to enhance efficiency. On a technical level, it is essential to unify technical standards for machine-readable formats, requiring that opt-out declarations adhere to specified data structures and are included in relevant EU standards catalogues. Existing AI model documentation forms could also be extended to serve as universal copyright declaration templates. Concurrently, a centralised copyright declarations portal should be established, guiding rights holders to use standardised language for submissions and ensuring interoperability with major technology platforms. Furthermore, support should be given to developing user-friendly data provenance tools to lower the technical barriers for transparency disclosures. To assist SMEs, a three-year transition period could be instituted, during which only general disclosures would be required, utilising funding from programmes such as “Digital Europe” to help them procure compliance tools. Judicially, the CJEU should be encouraged to clarify key concepts through preliminary rulings, drawing on Member State jurisprudence regarding the use of work excerpts to set explicit quantitative or proportional thresholds, thereby enhancing the practicality of judicial guidance.

Once the foundational rules are clarified, the governance approach should shift from a one-size-fits-all model to a tiered system based on the value-density of data. While the current legal framework primarily classifies and manages risks associated with AI applications, its key provisions do not genuinely follow a risk-based approach,Footnote ⁷¹ overlooking distinctions in the attributes of training data itself. For public domain and open-licence content, free circulation should be guaranteed. For vast quantities of low-originality, general data, statutory exceptions or implied licensing mechanisms could be introduced to reduce compliance costs. Regarding core data with high commercial value, carefully designed licensing and compensation mechanisms are necessary to protect rights holders’ interests, accompanied by stringent antitrust scrutiny to prevent data monopolisation. Such tiered governance can alleviate difficulties in accessing core data while achieving a more resilient balance between innovation efficiency and rights protection.

Regarding liability allocation, there is a need for innovative application of safe harbour principles. AI service providers should bear a reasonable duty of care concerning model outputs, including establishing content filtering mechanisms and infringement complaint channels. If a platform has fulfilled these obligations and responds promptly, liability limitations could apply to its use of training data and user-generated content. As a counterbalance, platforms should proactively disclose the basic types and source characteristics of their training data. This not only fulfills transparency obligations and builds industry trust, but also provides practical leads for rightsholders to safeguard their rights. At the same time, the implementation of this obligation must be tailored to specific contexts, maintaining a dynamic balance between technical costs and practical utility, ensuring that transparency requirements are both principled and reasonable as well as actionable.Footnote ⁷² Through these concrete pathways, a sustainable equilibrium incentivising technological innovation while protecting legitimate rights can be constructed.