“Generative AI And Copyright Global Policy And Legal Battles US vs EU Approaches”

I. Introduction: The Generative AI Challenge to Copyright Fundamentals:

1.1. Defining the Disruption: LLMs, GANs, and the Blurring of Creative Agency:

Think about generative AI—things like the massive text models, such as OpenAI's GPT series (Large Language Models, or LLMs), or those visual synthesis frameworks like Generative Adversarial Networks (GANs). They're doing more than just helping out; they’ve completely changed how we make content. These systems are fundamentally engineered to create novel outputs, spanning text, images, music, and video, by basically looking at vast amounts of existing data and figuring out the underlying statistical patterns ¹,.² This technology has rapidly spread and become deeply integrated into creative sectors, sometimes simply augmenting a task, but in many cases, outright replacing tasks traditionally performed by human artists, writers, and designers.

This technological acceleration presents a profound legal headache because it fundamentally blurs the line between works generated by a human and those produced by a machine. Traditional legal doctrine, particularly copyright law, is built on the singular premise of human intellectual effort and creative labour as the exclusive basis for protection. The capacity of an AI to autonomously produce outputs that convincingly

appear original necessitates a complete and urgent rethinking of established concepts of authorship, ownership, and liability. The core difficulty facing global regulatory systems, in my view, is that AI technology is evolving at a pace that far outstrips legislative capacity. Policy efforts, therefore, feel reactive, often attempting to fit cutting-edge 21st-century technological capabilities into 19th-century legal doctrines. This inherent time lag inevitably fosters the legal uncertainty we see everywhere and underscores the immediate necessity for adaptive legal frameworks.

1.2. The Doctrinal Crisis: Human Authorship as the Constitutional Cornerstone:

Okay, but what's the actual legal problem here? It all comes down to authorship. Copyright frameworks in both the U.S. and the EU are built on the idea that only humans can be authors. It's an old, fundamental rule. This entrenched principle explicitly excludes autonomous AI systems and other non-human entities from any recognition as legal authors or rights holders. The resulting legal vacuum concerning the ownership of works produced without significant human creative direction complicates both rights management and enforcement immensely.

Severity of copyright infringement even in the AI context is extremely complex. The core legal question is whether the process of training AI models, which usually involves mass reproducing copyrighted works at high volume without explicit authorization, is an unauthorized use or is allowed under existing exceptions. It creates a difficult bind: how do you defend the economic rights and market value of creators of content while fostering innovation and technological progress, which depend on access to massive, diverse training datasets? When an AI generates something truly independent, we end up with this ownership vacuum. Policymakers are caught in a tough spot: should they come up with brand new sui generis rights to reward the developers who built the model? Or should they just push these highly autonomous works straight into the public domain to prevent a handful of tech companies from essentially monopolizing culturally significant content? It’s a tricky balance.

1.3. Scope and Methodology of Comparative Analysis:

For this analysis, I'm focusing on the contrasting legal and policy approaches adopted by the United States and the European Union. These two jurisdictions are clearly at the forefront of generative AI regulation and jurisprudence, Their responses are shaped by fundamentally different legal traditions: the US operates under a common law system characterized by flexible doctrines like fair use and litigation-driven precedent, whereas the EU utilizes a civil law framework that emphasizes codified rights, prescriptive statutory exceptions, and centralized, rights-oriented regulatory oversight.

^{Our framework focuses on three key regulatory axes: the first is the dominant legal tool for permitting or excluding the mass use of copyrighted training data; the second is the legal attribution of authorship and ownership in the output generated by the AI derived from the copyrighted materials; and, third, emerging regulatory calls for mandatory transparency and accountability throughout the complex AI supply chain. Through comparison and contrast analysis of these approaches, the report analyses how differing national regulatory philosophies result in different balances between encouraging rapid technological progress and ensuring the effective protection of IP rights.}

II. Legal Frameworks Governing AI-Generated Works in the US:

2.1. Judicial Application of the Fair Use Doctrine in Training:

The U.S. legal system’s default answer to the problem of training AI is simply fair use (under 17 U.S.C. 107). This is a highly adaptable, factor-based defence that AI developers frequently cite to justify using copyrighted works—often without permission—to feed their models. The flexibility inherent in this doctrine is widely seen as critical to U.S. technological innovation, mainly because it minimizes the initial transaction costs and licensing overheads for foundational model development.

Judicial interpretations have centred on the factor of transformative use, assessing whether AI training is a productive, new use that extracts unprotectable patterns and knowledge, rather than merely replicating the original content in a competitive manner. In recent high-profile cases, such as the Anthropic Decision, courts appear to have emphasized that using legally acquired copyrighted books to train LLMs constitutes transformative fair use. The basic reasoning is that the AI systems learn underlying patterns to generate new, context-aware content, rather than directly competing with the original literary market, In effect, the U.S. reliance on this flexible defence acts as a de facto subsidy for AI innovation by minimizing initial licensing friction.

However, the application of fair use is not guaranteed, and the doctrine involves significant nuance. The Supreme Court precedent in Warhol dictates that transformation is a matter of degree, and it must be carefully measured against the other statutory factors, particularly commercially and the effect on the potential market for the original work. Furthermore, courts have signalled that the legality of the training process itself is paramount. Judicial guidance indicates that great weight is given to the provenance of the training data; if works are acquired unlawfully or through piracy, the fair use defence is likely to be rejected, irrespective of the model's transformative capacity. Such an emphasis on data’s legal origin creates an indirect point of regulatory leverage in the US. Although this flexibility undoubtedly promotes innovation, it also creates deep legal uncertainty and leaves both developers and rights holders vulnerable to expensive, drawn-out, and fact-intensive litigation.

2.2. The Human Authorship Requirement and the Limits of US Copyright:

^{Under the current U.S. copyright system, the need for human authorship remains a rigid requirement for protection, The Office has consistently reiterated this principle, explaining that copyright serves to protect the results of human intellectual effort – and it goes without saying that it does not consider the results generated by AI systems as a legal author.}

This stringent doctrinal requirement necessitates a precise delineation between works merely assisted by AI and those generated autonomously. The USCO has clarified that AI outputs are eligible for protection only where a human author has exercised sufficient creative control over the expressive elements. This includes human actions such as selection, arrangement, or modification of the output, or when the human-authored contribution is clearly perceptible within the AI-generated material. It's important to be clear: the USCO has explicitly said that just typing a prompt into an AI is not enough to claim authorship. It seems there can be no protection if the essential elements of authorship-the selection, arrangement, or expression-are conceived and entirely executed by the machine; it allows for a protection of AI-assisted productions where the human manipulates input to influence output but denies copyright to those autonomously generated. This attitude was bolstered in the District Court for the District of Columbia in Thaler v. Perlmutter (D.C. Cir. Feb. 9, 2024), which upheld the USCO’s denial of registration for an artwork where the authorship was attributed to an AI algorithm.

The USCO has maintained that existing copyright principles are sufficiently flexible to accommodate human-authored works that incorporate AI-generated material , concluding that major legislative changes regarding the copyrightability of AI outputs are not currently necessary. Nevertheless, the strict adherence to the human authorship requirement creates a palpable "copyright gap" for commercially valuable, highly autonomous AI outputs. Policy debate suggests addressing this gap through a two-tiered legal mechanism. This proposed framework would grant sui generis rights (a kind of non-copyright IP) to AI works generated with human contributions, while placing works generated entirely autonomously into the public domain. The existence of the "copyright gap," however, presents a risk of distorting creative incentives, potentially encouraging AI developers to seek only minimal human intervention—just enough to secure protection—rather than focusing on truly meaningful human-AI collaboration.

2.3. Emerging Policy Solutions: From DMCA Safe Harbour to the Proposed AI Harbour Model:

The limitations of the existing legal framework, particularly the Digital Millennium Copyright Act (DMCA), have become apparent. The DMCA’s safe harbour provisions, originally designed to shield passive online service providers, are ill-suited to generative AI systems, which actively ingest data and create dynamic outputs.

To address the unique risks associated with the AI supply chain, a fascinating new idea called the AI Harbour Model has been proposed. This model seeks to modernize the conditional-immunity bargain originally established by the DMCA by building an entirely new statutory section to manage infringement risks. The model is predicated on the idea of differentiated liability, justified by the fragmentation of responsibility across the complex AI development pipeline. Immunity is tied directly to the compliance of specific, role-relevant duties along the chain, promoting both efficiency and effective deterrence.

The AI Harbour Model segments responsibility across three principal roles, ensuring that compliance obligations are distributed to the entity best positioned to manage them :

Role	Function in Supply Chain	Role-Specific Duties for Immunity
Data Suppliers	Entities at the upstream edge, assembling or licensing massive datasets.	Provenance Disclosure and Transparency (Screening materials and disclosing licensing information early on to reduce downstream risk).
Model Developers	Entities designing and training the core architectures.	Dataset Curation, Memorization-Mitigation, and Watermarking (Integrating technical safeguards and content filters).
Model Deployers	Entities integrating models into consumer-facing services (final gatekeepers).	Dynamic Filtering, Complaint Handling, and Repeat-Infringer Policies (Quickly responding to specific output-based complaints).

Implementation of this model would require enacting a standalone statutory provision. Critically, the proposal mandates a transition toward administrative oversight via a specialized body: an

'AI Division' within the Copyright Office, This division would operate outside the traditional court system, performing essential functions such as certifying AI actors, conducting periodic audits (potentially through mandated self-assessment reports detailing compliance steps), and endorsing technical standards developed by industry co-regulation. This policy move suggests a critical shift in the U.S. approach: the traditionally litigation-centric jurisdiction is acknowledging that pure ex-post judicial remedies are insufficient to manage systemic risk in opaque AI models, pushing the U.S. toward a proactive, administrative oversight model that bears philosophical similarities to the European regulatory approach.

III. Legal Frameworks Governing AI-Generated Works in the EU:

3.1. Foundational EU Copyright Principles: Moral Rights and Human Creativity:

Over in the European Union, the philosophical starting point is quite different. Since their law is based on civil law principles, the framework is much more rights-oriented, placing immense value on the creator’s moral rights and the deep, personal connection an author has to their work. The EU framework rigorously adheres to the requirement of human creativity as the essential condition for copyright protection, generally denying legal status to works created solely by machine.

^{The EU focus on moral rights (namely the right of attribution and the right of integrity) certainly brings its own complications in the age of generative AI. When works are produced by AI systems which diminish human creative agency or imitate a creator’s style or unique expression, it gives rise to difficult issues as to how to safeguard the author’s personal and non-pecuniary interests. Legal and academic discussions continue to grapple with how to best apportion rights between developers, data owners, and users, and how to suitably incentivize innovation while preserving authors' economic and moral rights.}

3.2. EU Digital Single Market Directive and Text and Data Mining (TDM) Exceptions:

The European Union utilized the Directive on Copyright in the Digital Single Market (DSM Directive) to establish mandatory, codified exceptions for Text and Data Mining (TDM). This approach provides rule-based certainty, contrasting with the fluid nature of the U.S. fair use doctrine.

The DSM Directive distinguishes between TDM for non-commercial research purposes and TDM for commercial uses (Article 4), the latter being the core mechanism governing the training of commercial generative AI models. Commercial TDM is permitted, provided access to the content is lawful, but this use is strictly contingent on the condition that the right's holder has not expressly reserved the use of their works. This "opt-out mechanism" is a defining feature of the EU’s rights-first policy, granting content owners substantial control over whether their work is used for commercial AI training. It’s noteworthy that Article 4 was a late addition to the Directive, adopted before the current generation of GenAI had fully emerged, which may suggest concerns about its suitability for the contemporary technological landscape.

3.3. Practical Impediments and the Crisis of the Opt-Out Regime:

Despite its intent to provide clarity, the TDM opt-out mechanism has run into significant practical difficulties. Experts have confirmed that Article 4 of the CDSMD, not being developed specifically for the contemporary GenAI scenario, may be dysfunctional and unlikely to yield optimal results.

A major technical ambiguity surrounds the requirement that the reservation of rights be expressed in an "appropriate manner, such as machine-readable means" for content made publicly available online. This phrasing has created legal uncertainty regarding the technical validity of the opt-out. Although industry best practices such as robots.txt are preferable, case law, including recent decisions of German courts, seems to suggest that a clear, explicit opt-out formulated in natural language in the terms and conditions of a website may be accepted as a sufficiently machine-readable statement. The divergence in national legislation implementation in the Member States, in particular the fact that Italy does not provide details on machine-readable formats, is further complicating and fragmenting the whole scenario in the EU.

The consequence of all this complexity is a dangerously fragmented training data pool . This fragmentation may very well hold back AI development in the EU, putting European developers at a competitive disadvantage globally. Frankly, the EU's rights-first approach, which was intended to provide certainty, seems to be accidentally creating significant friction for accessing data, especially when compared to the U.S. framework which grants broad, if litigious, data access under the fair use doctrine. This instability has driven growing momentum within the European Parliament to review or overhaul the framework, potentially emphasizing clear consent, remuneration, and greater transparency as more robust mechanisms than the current burdensome opt-out regime.

3.4. Regulatory Intervention: The EU AI Act as a Compliance Layer:

Recognizing these weaknesses in the TDM regime, the EU stepped in with the AI Act. This legislation sets up what I would call a layered governance model , addressing the limitations of the DSM Directive. The AI Act imposes specific, mandatory transparency requirements that overlay and actively enforce existing EU copyright obligations, particularly for General-Purpose AI (GPAI) models.

Providers of GPAI models therefore have a key duty under the AI Act to comply with EU copyright law and, in particular, to make instead of the full copyrighted data used for training a sufficiently detailed summary available. This should increase accountability in a way that would allow right holders and regulators to ascertain whether TDM exceptions and copyright in general are being followed. It also includes an obligation to disclose for deployers, stating that content that is artificially generated – especially deepfakes – should be clearly labelled as such, with exceptions for artistic works when such disclosures would affect the consumption. The EU AI Act is unique in its willingness to use horizontal risk and safety regulation to enforce specific IP obligations. This proactive regulatory philosophy aims to anticipate and mitigate legal conflicts through administrative mandates, providing a sharp contrast to the U.S. reliance on ex post judicial resolution.

IV. Comparative Analysis: US vs EU Approaches to AI Training Data:

4.1. The Axis of Conflict: Flexibility (US) vs. Certainty (EU):

The core difference is this: the US leans hard into flexibility. They allow for quick technological moves through case-by-case fair use decisions, accepting legal uncertainty as just part of the price of progress . The system, therefore, often feels reactive, with policy evolving primarily through the results of high-stakes litigation.

Conversely, the EU prioritizes legal certainty through codified statutory exceptions and comprehensive, centralized regulatory mandates. While this approach provides rights holders with greater control via clear mechanisms like the opt-out, it risks restricting data accessibility necessary for cutting-edge innovation. The U.S. framework effectively subsidizes innovation speed due to low initial data acquisition friction under fair use; the EU framework, by contrast, prioritizes lawful innovation and compliance, potentially leading to safer, more rights-respecting models, even if the pace of development may be constrained by data friction imposed by the opt-out mechanism.

4.2. Licensing and Remuneration Models:

When it comes to paying creators, the economic models also diverge. In the U.S., the system is fundamentally market-driven and enforcement-centric. Licensing is voluntary and heavily influenced by the constant threat and specific outcomes of fair use litigation, forcing rights holders to bear a high burden of enforcement. This structure often favors the largest rights holders who can afford protracted legal battles .

The European Union, structurally, seems much better suited to move toward implementing collective licensing and fair remuneration schemes. Think of Copyright Management Organizations (CMOs) stepping in to manage licensing for countless rights holders at once. This streamlines the process for developers trying to source data and ensures compensation flows back to creators. Analysis may suggest that the EU's collective licensing model, if successfully standardized, offers a more sustainable long-term solution for managing mass data access than the potentially volatile, litigation-driven U.S. system.

4.3. Governing the Output: Ownership, Liability, and the Tiered Protection Debate:

A doctrinal consensus exists between the two jurisdictions: human authorship is indispensable, and copyright protection is denied to works generated solely by autonomous AI systems, The active policy debate centers on how to allocate and protect the human investment required to deploy sophisticated AI tools that generate valuable content.

To resolve the resulting legal ambiguity, academic proposals advocate for a two-tiered protection system. This system would assign sui generis rights—a non-copyright form of intellectual property—to AI-assisted works that demonstrate significant human creative contribution, thereby incentivizing investment, while committing genuinely autonomous outputs to the public domain.

In terms of liability assignment, both the U.S. and EU recognize that responsibility is fragmented across the AI supply chain. The U.S. proposal for the AI Harbour model wants to establish clear, role-based responsibility overseen by an administrative body. The EU, relying on its regulatory tradition, achieves differentiated accountability by coupling its existing copyright liability rules with the proactive transparency and enforcement mandates established under the AI Act. These parallel efforts demonstrate a shared policy realization that liability management requires clear, distributed standards enforced beyond reliance solely on judicial interpretation.

V. Litigation as Policy: Landmark Legal Battles Defining the Landscape:

5.1. The Battle for the Pixels: The Getty Images v. Stability AI Saga:

The ongoing fight between Getty Images and Stability AI is a classic case that essentially defines the conflict over training image models, Getty is accusing Stability of infringing millions of their copyrighted photos—plus all the associated captions and metadata—to train the Stable Diffusion generative model. The core legal question, of course, is whether this mass ingestion constitutes copyright infringement or falls within the scope of fair use .

It's not just about copyright. They also threw in serious allegations of trademark infringement. Why? Because Getty claims the AI model can sometimes reproduce their famous watermarks in the generated images. If that’s true, it really calls into question Stability’s entire argument that the training process is "transformative" and non-infringing.

Crucially, the parallel U.K. proceedings underscored the immense opacity barrier facing rights holders. Getty ultimately dropped its primary infringement claims related to copying during the development stage, proceeding only on secondary infringement. This experience highlights the substantial burden of proof placed upon rights holders attempting to trace specific infringing actions within the technically opaque, multi-stage, global training pipelines of large models. The difficulty of obtaining ex post judicial remedies in such complex cases provides compelling support for the political and technological necessity of the U.S. AI Harbour proposal, which mandates ex ante transparency and accountability to manage systemic risk.

5.2. The Writers' Resistance: Authors Guild and the Future of Text Generation:

The Authors Guild v. OpenAI litigation is a seminal case challenging the transformative use defence in the context of Large Language Models (LLMs) and literary works. OpenAI's defence rests on technical arguments regarding tokenization and pattern learning, asserting that the training process extracts only unprotectable information, fundamentally transforming the copyrighted content.

The problem, and the main counter-argument, is the simple fact that some LLMs, including ChatGPT, can be made to spit out verbatim chunks of their training data under certain conditions. This ability to reproduce source text seriously complicates the legal argument that the use is purely non-competitive and transformative, compelling courts to weigh the factor of market harm more heavily. The documented risk of memorization and verbatim reproduction shifts the legal focus from the legality of the input (training data acquisition) to the potential infringement in the output (the content generated). I think this means that future legal standards must be highly model-specific; what constitutes "transformative" for a language model prone to text memorization may be different from the standard applied to an image model.

5.3. Implications of Emerging Music and Code Copyright Suits:

Legal uncertainty is rapidly spreading across all creative sectors, demonstrating the systemic nature of the challenge. The music industry is speaking out, with groups like the International Confederation of Music Publishers (ICMP) describing major AI companies’ actions as “wilful, commercial-scale copyright infringement” in relation to the unlicensed use of entire music catalogues for training purposes.

Among recent actions such as the Recording Industry Association of America (RIAA) against AI music generators like Suno and Udio, they aim to establish a case law of infringement for when AI models create music that mimics the style, voices or melodies of known artists. The unprecedented surge of these lawsuits across visual arts, literature, music and code demonstrates that the legal regime right now is carving through sectorial precedent, with the need to provide more coherent and harmonized legal certainty so that all stakeholders can have stability being more urgent than ever.

VI. Data Governance, Ethics, and Accountability in AI Development:

6.1. Data Provenance and Ethical Curation Practices:

Any high-performing AI model, by its nature, is heavily dependent on having access to colossal, varied datasets. Because of this, making sure that data was lawfully acquired—having solid data provenance—is absolutely crucial. The precise copyright status of these datasets and whether they align with exceptions often remains unclear, significantly complicating accountability and rights enforcement.

Model developers and data curators are consequently tasked with implementing technical safeguards to meet ethical and legal compliance standards. Techniques such as memorization mitigation—engineering methods to limit the AI’s ability to reproduce specific training examples verbatim—and the implementation of watermarking or traceable markers on generated outputs are becoming essential technical tools for compliance and risk reduction. This shift means that compliance has moved beyond a purely legal requirement into a complex technical engineering challenge. Failure to implement robust mitigation measures may, in future litigation, be interpreted as a failure to satisfy a developer’s 'duty of care' during the model training phase.

6.2. Require Transparency: Reporting and the Public Trust:

Internationally, legislative proposals are increasingly focused on requiring transparency in the use of AI training data, as it is needed for the public trust and appropriate regulatory supervision. The EU AI Act in particular emerges as a trailblazing regulatory project, requiring GPAI model developers to provide training data summaries to aid in monitoring and rights enforcement. This level of transparency is needed to address the information asymmetry between large, monolithic AI developers and fragmented, individual rights holders, thus raising the latter’s ability to meaningfully enforce their rights or negotiate compensation.

In parallel policy debates in the U.S., including the U.S. Copyright Office’s reports on digital replicas (deepfakes) and recommendations for new federal rights to control unauthorized use of voice, image, and likeness (e.g., the NO AI Fakes Act), there is a growing blurring of IP policy with ethical governance. This is emphasizing a global approach to ensuring transparency in the management of ethical risks, protection of identity and fraud prevention management in the era of enhanced digital replication.

6.3. Advanced Accountability Models: Certification and Auditing (The Role of the USCO AI Division):

Assigning liability and accountability in the complex AI ecosystem, involving data suppliers, developers, deployers, and users, remains a formidable challenge. Effective accountability requires clear, differentiated standards for compliance.

The U.S. AI Harbour proposal offers a scalable solution by suggesting the creation of a specialized 'AI Division' within the Copyright Office, This administrative body would manage regulatory oversight, performing functions such as actor certification and periodic auditing (potentially via mandated self-assessment reports). By institutionalizing this administrative oversight, the proposal acknowledges that the technical complexity of AI necessitates specialized, non-judicial supervision, akin to successful regulatory models used by agencies such as the FTC or SEC, to handle highly technical copyright compliance issues efficiently.

6.4. Ethical Obligations and Moral Rights in the Age of Replication:

^{In addition to legal and economic debates, generative AI presents profound ethical and normative challenges around, among other issues, moral rights. Policy needs to achieve a balance between fostering technological innovation and ensuring ethical accountability and equitable treatment of the original creators, whose work serves as the input to AI models.}

^{Maintaining authors' moral rights—such as attribution and integrity—is a key ethical concern. If an AI produces a generic work, that's one thing, but when models are trained to replicate the voice or style of a individual human creator, that's a different matter. This seems to be a personal attack on the creator's own work and it's borderline exploitative unless they get paid or given any control. Long-term societal acceptance for GenAI is conditioned on not only a model for how creators can be compensated economically (royalties), but also on the respect for the non-pecuniary rights of creators the right to control association with and integrity of their style of creativity. Clear legislative definitions of these ethical duties is paramount in retaining public trust and eliminating exploitation in the creative industry.}

VII. Exceptions, Fair Use, and Opt-Out Mechanisms:

7.1. Fair Use Doctrine in the US Context:

In the U.S., fair use, for all its faults, has been central to AI training. It is an adaptable framework that tries to balance public access to knowledge with the private rights of creators. Its focus on transformative usage means that AI models can be trained on copyrighted material in a way that the use is argued to be productive and not market substitutive, suggesting a lower risk of initial infringement.

But the main problem with this loose standard is that it creates a lot of legal uncertainty. Since fair use is a fact intensive inquiry and a determination of the issue does not apply broadly, AI developers are currently unable to confidently predict the outcomes of legal disputes, resulting in an active, if legally unstable, environment.

7.2. Text and Data Mining (TDM) Exceptions in the EU:

The EU’s solution in the DSM Directive is radically different: it introduces specific TDM exceptions, aiming to foster more rule-based certainty, At the centre of the EU's design is the mandatory Rights holder opt-out Claus for commercial TDM, designed to provide copyright owners with clear, statutory control whether their work can be used for AI training.

The enforcement of this opt-out mechanism, for example, has led to strong opposition, for it might deter access to data and development of AI. Moreover, the differing implementation of the 'machine-readable means' requirement by the Member States creates additional challenges for the developing AI industry and will lead to further fragmentation of the European internal market in this field.

7.3. Proposed Hybrid and Alternative Models:

The problems with both the hyper-flexible U.S. model and the rigid EU opt-out approach have really increased interest in hybrid models. These frameworks are designed to integrate the best aspects of both models: the clarity and administrative guidance of the EU model with the flexible, fact-based analysis appropriate for an application involving rapidly evolving technologies, as lauded by the U.S.

Such submissions often call for differentiated treatment by use, require remuneration schemes (such as collective licensing), with further technical transparency obligations. Such blended-type models seem to flnd a good compromise, and might be the right way to establish balanced, legally solid regimes that reconcile the competing policy concerns to promote radical AI innovation on the one hand and maintain strong and meaningful rights protection on the other.

VIII. Conclusion: Shaping a Sustainable Ecosystem for Creativity and Innovation:

What generative AI has really shown us is the deep structural fault lines in global copyright law. The U.S. is clearly prioritizing speed and innovation, leaning heavily on the uncertain fair use defence, which inevitably leads to a ton of expensive, high-stakes court cases over training data. The European Union, on the other hand, puts a premium on legal certainty and creator control through codified rules, but that runs the risk of creating too much friction and slowing down their own AI development pace due to the technical complexities of the TDM opt-out mechanism.

I think we're seeing a necessary convergence in policy responses, though: neither a system of only reactive litigation nor one built on overly prescriptive, rigid statutory restrictions seems to be working perfectly. The U.S. proposal for a proactive, administrative AI Harbour Model and the EU’s use of the AI Act to enforce GPAI transparency both acknowledge the imperative of differentiated accountability and ex ante transparency to manage the opacity inherent in large generative models.

Way Persistent legal uncertainties still exist, in particular concerning authorship for works generated by human-AI co-creation and the development of scalable, global liability models. The line of doctrine is still clear: human creativity is a must for copyright protection. Hence, the future of policy has to be about how to develop nuanced answers – including perhaps tiered (and even sui generis) rights that can reward human investment and contribution while preserving the overall vitality of the public sphere.

The adaption of (legal) frameworks that allow for transparency, foster the development of efficient collective licensing models as means for compensation and reduce the (often very costly) individual litigations might be the way forward to develop a truly sustainable system of culture production. In addition, active international cooperation under the aegis of such organizations as WIPO is indispensable for ensuring the unification of conflicting legal systems and effective transnational execution. By merging the U.S. principle of transformative use with the EU’s dedication to administrative transparency—let’s call it a Hybrid Compliance Framework—policymakers can foster an environment where generative AI can thrive in a world grounded by a strong, rights-protecting intellectual property system.

Reflective Questions:

Question: How does the U.S. reliance on flexible fair use for AI training create both a benefit for innovators and a persistent risk for developers?

Answer: Fair use aids rapid innovation by lowering initial licensing friction. However, transformation is highly nuanced. Courts heavily weigh the  provenance of training data; if works are acquired through piracy, the defence is rejected, leaving developers exposed to high litigation risk despite technical transformation.

Question: The EU attempted to ensure legal certainty for AI training via the TDM opt-out. Why has this rule-based approach been called "dysfunctional"?

Answer: The TDM rule failed to clearly define "machine-readable means," causing fragmented implementation across member states. This ambiguity creates structural data friction, limiting the total training pool and potentially hindering EU-based AI competitiveness, despite the goal of providing certainty.  

Question: What major philosophical shift does the U.S. AI Harbour Model signal by proposing an 'AI Division' within the Copyright Office?

Answer: It shifts the U.S. from purely ex post judicial enforcement to ex ante administrative oversight. The 'AI Division' would proactively manage systemic risk by auditing and certifying actors based on compliance with specific technical duties (like memorization mitigation), addressing the black-box nature of models.

Disclaimer: The content shared in this blog is intended solely for general informational and educational purposes. It provides only a basic understanding of the subject and should not be considered as professional legal advice. For specific guidance or in-depth legal assistance, readers are strongly advised to consult a qualified legal professional.