Beyond the prompt: reconciling training data ingestion and music copyright in the generative AI era

Oblikovanje medijske budućnosti u digitalnom okruženju

1Faculty of Management, Sremski Karlovci, Serbia ,

The global music industry is currently undergoing unprecedented regulatory restructuring driven by the rapid commercialization of generative artificial intelligence (GenAI). The music business is no stranger to disruption, from the transition to digital formats, through peer-to-peer file sharing that decimated revenues at the turn of the millennium, to streaming services that moved the industry from an ownership model to a subscription one. However, the current disruption is fundamentally different: it does not merely alter the distribution channels and consumption habits. Rather, it changes the core mechanics of music creation, challenging our existing definitions of authorship and the foundational architecture of Western copyright laws. The tension between computers and authorship has existed since the 1960s when scholars and legal practitioners first began analyzing the implications of computers on copyright frameworks (Miller, 1993), but in recent years, it has escalated from theoretical debates into real regulatory actions and legal cases.

The Mechanics of Generation: Training Data and the "Black Box" Problem

The catalyst for these legal cases lies in the fundamental architecture of the technology itself. To train AI models capable of high-quality, realistic music generation, developers must inject vast quantities of published sound recordings and musical compositions into their models, a substantial amount of which remain under copyright protection rather than in the public domain. However, these models do not disclose the specific data points referenced when generating outputs from user prompts, nor do they

© 2026 by the authors. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license .

reveal the protected works from which those data points originate. This issue, in which only the model's outputs are visible while the training inputs and the underlying generative mechanics remain obscure, is commonly referred to as the AI "black box" problem.

The algorithmic black box introduces a plethora of complications. While the ultimate severity of these issues will depend on emerging legislation and judicial rulings in ongoing legal cases, a few immediate evidentiary problems will almost certainly require solving. Primarily, a traditional copyright infringement claim requires the plaintiff to prove two key elements: that the defendant had access to the copyrighted material, and that there is a "substantial similarity" between the protected work and the alleged infringement. However, without full transparency from AI proprietors regarding their training datasets, proving the element of access becomes difficult. Furthermore, while the "substantial similarity" test functions well for traditional infringement such as the direct, unauthorized use of a specific melody or lyric, it becomes highly ambiguous in the context of generative AI. Because these models utilize deep neural networks (not unlike the human brain) and pattern recognition, they often generate outputs that are arguably entirely new in their melody and lyrics, even if protected works were used as the foundation for their creation.

Another, yet equally important, issue is intent: how does a plaintiff prove the "intent" of an artificial intelligence model to infringe upon a copyright? One might argue that intent is implied at the training stage, given that the primary purpose of ingesting protected works is to train the model to replicate their stylistic and structural characteristics. Legal scholars have argued that foundational doctrines of U.S. law (specifically intent and causation) are fundamentally ill-equipped for adjudicating on copyright infringement by artificial intelligence models (Bathaee, 2018).

The Authorship Problem

Chris Cooke of CMU writes that the debate surrounding the copyrightability of entirely AI-generated works is framed around two core issues: whether existing copyright laws provide the framework and wording necessary for protection, and fundamentally, whether such works warrant protection at all (Cooke, 2026). However, if lawmakers and the judiciary at any point decide that these works warrant protection, an important third question emerges: who will hold the copyright? Would the rightsholder be the prompt provider (the user), the proprietary owner of the model (the developer), the artists and songwriters whose works were used as the training data, or a combination of all of the above?

Interestingly, there is a theoretical fourth option: the author could be the AI itself (Caldwell, 2023). Drawing a parallel from patent law, Abbott (2016) suggests that AI should take precedence over the

HcclkV © 2026 by the authors. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license .

user in claims of inventorship, arguing that “simply providing a computer with a task and starting materials would not make a human an inventor”. If this logic is translated into copyright law (treating AI models as autonomous authors of their creative works) it could introduce significant logistical hurdles. For example, if an AI is a legitimate copyright holder, how is the financial remuneration distributed when these works are exploited? Would royalties flow back to the developers, the prompt providers, or the original training data rightsholders? If an AI has no financial needs or incentives, the fundamental economic purpose of granting it copyright ownership is questionable.

On the other hand, there are valid arguments against granting AI the possibility to be the author. Craig and Kerr (2019) contend that granting authorship to an AI is a fundamental mistake rooted in the romanticization of artificial intelligence and a misunderstanding of what authorship is at its core. They argue that the debate should not focus on the technological sophistication of the models (what the AI can do) but rather on the ontological nature of authorship itself, emphasizing what an author must inherently be: a human with social relations.

Bifurcated Rights: The Dual Copyright Dilemma in Generative Audio

While philosophical debates regarding the nature of authorship mostly favor the view that copyright is inherently tied to human conduct, applying this principle to the music industry requires navigating highly specific structural complexities. Because the music industry fundamentally relies on a dual copyright structure (one for the underlying musical composition, in the US often referred to as the PA, or Performing Arts copyright; and one for the sound recording, referred to as the SR copyright) any future framework extending copyright to AI-generated works must reconcile this dichotomy. Regardless of which entity might ultimately hold these rights, two primary options emerge: (1) adapting the existing dual framework to establish distinct rights for both the AI-generated composition and the sound recording, or (2) developing a new, sui generis copyright structure specifically tailored to machinegenerated works.

Under current U.S. law, a sound recording generated entirely by AI is ineligible for copyright protection due to the strict requirement for human authorship (U.S. Copyright Office, 2021, Section 306). Both the U.S. Copyright Office and European Union legal frameworks maintain that an AI cannot own or be granted a copyright. The UK, on the other hand, offers protection for computer-generated works lacking a traditional human author, but it does not grant ownership to the machine itself; rather, the law awards the copyright to the human or entity that made the necessary arrangements for the work's creation (Copyright, Designs and Patents Act, 1988, § 9(3)). In a scenario where a user provides original lyrics but an AI independently creates the musical arrangement and sound recording, the resulting copyright is bifurcated. The human-authored lyrics are fully protected by copyright, but the AI-generated sound recording and the melody are ineligible for protection and fall into the public domain as soon as they are created.

Determining this allocation of rights is inherently tied to a more complex technical challenge: identifying exactly which underlying copyrighted works informed a specific AI-generated output. Because tracing these outputs back to the original creators whose works were used in the training process is obscured by the "black box" nature of deep learning algorithms, the music industry’s strategy has largely moved away from claiming ownership over AI-generated outputs. Instead, rightsholders are targeting the foundational mechanics of the models themselves: the input phase. If the developers arguably cannot build commercially viable generative models without ingesting huge datasets of protected works during training, the fundamental legal question shifts from who owns the rights to

HcclkV © 2026 by the authors. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license .

generated songs to whether the AI companies had the right to train their models on copyrighted material in the first place. This conflict over unauthorized ingestion is the core of the current industry lawsuits.

Litigation and the "Fair Use" Defense

AI music generation platforms such as Suno and Udio have acknowledged their use of unlicensed, copyrighted music in their training datasets (Tencer, 2024). However, in response to litigation brought by independent artists (Justice v. Suno, Inc., 2025), the companies argued that the claims "fail as a matter of law" because their models do not actually sample any copyrighted audio. Controversially, Suno and Udio have asserted that ingesting copyrighted works for model training constitutes "fair use" under U.S. copyright law (UMG Recordings v. Suno, Inc., 2024). While the debate persists over whether AI-generated music is fundamentally transformative or derivative, the use is undeniably commercial as long as these platforms operate for profit. A central battleground within this fair use debate is the unauthorized ingestion of lyrics. Artists and songwriters are increasingly demanding the right to opt out of having their intellectual property utilized as training data. For instance, in March 2026, a coalition of independent songwriters and performers filed a class-action lawsuit against Google regarding its music generation model, Lyria 3 (Kogon v. Google, LLC, 2026). The plaintiffs alleged that the model was trained on approximately 370,000 hours of music without the explicit consent of the creators. They further contended that Google excluded independent artists from licensing negotiations held with major labels and publishers, which reportedly took place only after the model's training was already complete. Similarly, in late 2025, the Munich Regional Court in Germany ruled in favor of the performing rights organization GEMA, determining that OpenAI violated copyright laws by utilizing and reproducing copyrighted lyrics without securing the necessary licenses (Stassen, 2025).

An important factor in assessing a fair use defense is the potential market harm to the original copyright owner. Given that AI-generated tracks can serve as direct market substitutes for the underlying copyrighted works, there is strong presumption that platforms like Suno and Udio actively hinder the original rightsholders' earning potential. Asserting fair use remains a highly contentious defense when unlicensed, copyrighted works are used to develop commercial products that directly compete with the original creators. However, there are counterarguments supporting the unlicensed training of AI models under the purview of fair use. Torrance and Tomlinson (2023) contend that fair use should be recognized as a valid defense for AI training on copyrighted materials because it constitutes a transformative and non-consumptive process; a framework they call "fair training". They argue that AI learns in a manner akin to humans, by being exposed to a large amount of data. Because AI does not merely copy the original content, but rather generates significantly different material based on learned patterns when compared to any single protected work within its dataset, the process is viewed as transformative. Furthermore, since the training ingestion leaves the original data and works unaltered, the authors classify the process as non-consumptive.

However, the relevance of this non-consumptive categorization is highly questionable given that these AI models are inherently commercial products. They arguably capture market share from the original rightsholders, undermining copyright frameworks that were designed not only to prevent intellectual property theft, but also to incentivize artistic creation by allowing creators to dedicate themselves to their craft without being forced into other means of employment (frequently referred to as “day jobs”) to survive. If AI-generated music comes to dominate the market, it raises critical questions regarding the future of AI training. Specifically, how close are we to (1) entering a recursive AI training loop, often referred to as model collapse (Shumailov et al., 2024), that merely perpetuates existing

HcclkV © 2026 by the authors. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license .

patterns instead of generating genuinely novel works; while simultaneously (2) diminishing the number of human creators who can afford to produce truly unique art because AI has severely diluted and devalued the market?

Torrance and Tomlinson go on to conclude that “fair training becomes a necessary concept for the democratization and continued development of AI. The “fair training” exception will balance the rights of copyright owners with the AI’s ability to learn and grow” (2023, p. 10). However, they fail to adequately explain how "fair training" balances the rights of copyright owners, offering no real solution to the aforementioned market dilution. Generative AI models introduce unprecedented competition at the core level of creation, effectively driving the price of musical generation down to levels that no human writer or artist can match while sustaining a living wage. The authors appear to rely on semantically redefining AI's ingestion of copyrighted works as inherently non-infringing, an insubstantial argument that overtly prioritizes the unhindered development of AI completely disregarding creator protections.

As the music and broader arts industries navigate this profoundly disruptive phase, it is imperative to approach these issues with a higher degree of nuance than Torrance and Tomlinson provide. We must find a way to balance the beneficial development of generative AI models while also making sure that creative industries are not irreparably damaged and that the arts and culture sectors continue to thrive. As cultural economist David Throsby establishes in his book Economics and Culture (2001), the cultural capital is a foundational asset that generates both economic wealth and cultural identity. Depreciation of cultural capital cannot be easily reversed by substituting it with other forms of capital.

An Antithetical Legal Picture: The US, EU, and UK Approaches

Translating this necessary balance into actionable policy, however, has proven difficult. While Western copyright laws within the context of music are fundamentally premised on protecting both the underlying composition and the sound recording, their application still varies significantly across jurisdictions. Interpretations of what constitutes protected material differ significantly based on specific statutory language and judicial precedent. The emergence of generative AI has introduced deep uncertainty, raising questions not only about how existing frameworks will be interpreted, but also exposing regulatory vacuums that will likely require legislative amendment. The regulatory and judicial responses to this paradigm shift remain geographically fragmented, which creates a highly unstable environment for global licensing and intellectual property enforcement, particularly because the global nature of modern digital technologies transcends clear jurisdictional boundaries (Johnson & Post, 1996; Svantesson, 2017).

The United States (US), the European Union (EU), and the United Kingdom (UK) are managing the collision between latest algorithmic innovation and the protection of creative rights very differently. The United States has defaulted to an ex-post approach, mainly deciding not to intervene until litigations are started, relying on the judiciary to interpret the often-debated “fair use” doctrine. On the other hand, the European Union has been more proactive, adopting an ex-ante regulatory strategy aiming to enforce fundamental rights, transparency, and statutory mechanisms such as the AI Act. The UK occupies a conflicted middle ground, attempting to foster a competitive, somewhat deregulated technology market while simultaneously seeking to appease its large and influential creative sector. Ultimately, as AI becomes increasingly better at mimicking human judgments and decision-making within creative fields, scholars emphasize that our current legal frameworks will likely require fundamental revisions (Mammen et al., 2024).

In the US, the protection of creative works is inherently tied to the presence of human creativity.

HcclkV © 2026 by the authors. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license .

According to the Copyright Act of 1976, the rights are reserved for “original works of authorship”, a standard that has consistently been interpreted by courts and the US Copyright Office to require a human creator. A recent decision by the U.S. Court of Appeals for the D.C. Circuit confirmed this in Thaler v. Perlmutter (2025), with the Supreme Court subsequently denying review of the case in early 2026, and affirming that copyright law protects only works created by a human being (in this instance, Dr. Thaler unsuccessfully attempted to register visual art autonomously generated by his AI system, listing the machine as the sole author). Contemporary discourse suggests that creators will increasingly need to maintain rigorous documentation and evidentiary records of their workflow to substantiate human authorship, such as preserving input data for subsequent analysis or submitting iterative drafts of their work to prove the final output was not predominantly AI-generated (Fritz, 2025; Hugenholtz, 2024).

On the other hand, while the EU approach also mandates a human creator for copyright protection, it places a significantly heavier emphasis on the ingestion phase, specifically on the legality of the training data used to develop the AI models. Rather than mostly deliberating on the copyrightability of the AI generated outputs, EU policymakers and courts are increasingly viewing the legal status of the output as inseparably linked to the legality of the input data. According to the EU Parliament report on copyright and generative artificial intelligence, failing to comply with rightsholder opt-outs, utilizing pirated material for training, or failing to seek and obtain necessary licenses all constitute a violation of creators’ fundamental rights (Voss, 2026). Furthermore, the EU has assessed that general-purpose AI models placed on the EU market must comply with EU copyright laws, regardless of the jurisdiction where the AI training physically took place or the specific copyright statutes governing that location.

The UK briefly opted for a third variant approach in 2022, proposing a broad text and data mining (TDM) exception for any purpose (including commercial use) with no opt-out provision for rightsholders at all (Clark, 2026). This proposal was abandoned less than a year later following widespread backlash from the creative sector citing concerns with potential revenue loss.

However, between late 2024 and early 2025, the UK government launched a comprehensive consultation to reassess its regulatory approach to AI training and TDM. The consultation explored four avenues: maintaining the status quo (Option 0), strengthening copyright to require licensing in all cases (Option 1), introducing a broad TDM exception without an opt-out (Option 2), and the government’s then-preferred route: a TDM exception accompanied by a rightsholder opt-out mechanism (UK Copyright and AI Report, 2026). Ultimately, this "preferred choice" was officially abandoned in March 2026 after a landslide 81% vote favoring a strict licensing approach over TDM exceptions. Nevertheless, UK policymakers remain caught in a delicate balancing act. While the creative industries currently dominate in Gross Value Added (GVA), the government recognizes the huge projected growth of the AI sector and remains reluctant to implement restrictive copyright measures that might deter AI proprietors from entering or investing into the UK market.

Regardless of the specific regulatory trajectories, the remainder of 2026 will undoubtedly serve as a critical stage, establishing key milestones for how these three major jurisdictions govern the intersection of generative AI and intellectual property in the future. These developments will be of more than just local and immediate concern, as the legislation enacted in the US, EU, and UK will do more than just shape domestic markets, as it will probably be setting a foundational global precedent that the rest of the world is likely to follow.

Proposing Scalable Solutions: Attribution and Blanket Licensing

Given the fragmented and unpredictable nature of these global regulatory approaches, waiting for a

HcclkV © 2026 by the authors. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license .

unified legal consensus may leave rightsholders exposed to continued market dilution. Because the borderless nature of generative AI makes localized copyright enforcement difficult, the industry should explore proactive, scalable technological solutions alongside legislative efforts. Solving the legal, jurisdiction-transcending disputes requires first solving the technical problem of the algorithmic black box. To address this, a team of researchers from Sony published a paper detailing a process for Training Data Attribution (TDA), which aims to determine the original works used in the model’s output (Choi et al., 2025). A clever technique they use is machine unlearning, where by forcing the model to forget specific data points, researchers can deduce whether that data was used in a generated output by observing any resulting degradation in quality; if the output gets worse or less accurate, it means that the omitted work was a source for that specific generation. Should this technology prove to be viable and widely available for commercial use, it could provide creators and copyright holders with a quantifiable mechanism to substantiate their demands for licensing fees and remuneration for the use of their protected works.

However, while TDA provides the technical capability to pinpoint ownership, attribution alone is only half the solution. Even with a perfect system for identifying the underlying rights, strictly enforcing traditional reproduction rights based on these findings would ultimately stifle the technology's utility. It is practically impossible for companies that own AI models to clear individual micro-licenses for every single user prompt or a single copyright, whether for compositions or fully rendered audio (Stout et al., 2026). While the AI generation process structurally resembles a synchronization license, where underlying works are integrated into a newly generated media output, the sheer volume of daily user prompts renders traditional, individual sync-like negotiations completely unfeasible. The industry needs a scalable solution that can work for both the copyright holders and the AI model’s proprietors: a new type of comprehensive blanket licensing framework (administered much like the performing and mechanical rights managed by PROs and the MLC), that legally encompasses both the sound recordings (SR) and musical works (PA) without the bottleneck of frequent, individual clearance.

A logical blueprint for this could be the Mechanical Licensing Collective (MLC), which solved a similar logistical bottleneck for digital music streaming. However, rather than adopting the controversial pro-rata model widely used in on-demand music streaming, an AI collective could distribute royalties using a user-centric payment system. Instead of pooling together all subscription revenue which disproportionately rewards top rightsholders regardless of an individual AI user’s actual prompts, a usercentric approach isolates the licensing fraction of a specific user’s AI subscription. Working in tandem with Training Data Attribution (TDA) technologies, that fraction could be divided exclusively among the specific copyright holders whose underlying works directly informed that individual user's generations during the billing period (for a foundational industry analysis comparing user-centric and pro-rata revenue models, see Muikku, 2017).

This framework could create a practical middle ground: AI developers secure the blanket clearance they need to keep their interfaces useful; songwriters, artists, and other copyright holders are compensated fairly based on actual demand; and users ensure their subscription fees are allocated only to the creators whose works informed their generated outputs.

An alternative framework requiring a more proactive stance from rights holders would be the implementation of an opt-in, closed-universe database (Epstein, 2025). This system would aim to create a monetizable ecosystem for rightsholders whose intellectual property is used in generative AI music models. Under this model, composers, lyricists, recording artists, or record companies could voluntarily submit their protected works into a controlled database which would constitute an authorized pool of material for AI to draw from. This closed-universe solution could mitigate the systemic issue of

the Creative Commons Attribution (CC BY) license .

unauthorized data ingestion and addresses the pervasive black box problem currently destabilizing the industry. Moreover, the proposal calls for a transparent mechanism for Training Data Attribution (TDA) via an “ingredient label” conceptually analogous to nutritional labels found on commercial comestibles. This attribution model would systematically delineate the specific underlying works used in a given AI output, detailing both the proportionate percentage of usage and the nature of the protected material (e.g., musical composition versus sound recording). Another benefit of this proposed system is the likely ability to facilitate automated royalty distributions directly proportional to the TDA metrics, mitigating the potentially costly legal disputes stemming from copyright infringement.

At the time of this writing, one of the most recent implementations of such a system involves three new features (Variations, Craft, and Magic Fit) introduced by the sample-provider platform Splice. These features allow users to modify existing samples from the Splice catalog, adjusting parameters such as structure, key, and tempo while preserving the sonic characteristics of the original sound (Stassen, 2026). The company claims that the tools track every sample utilized in their output, maintaining attribution and ensuring remuneration for the original creators each time their sounds are sourced or downloaded (Dredge, 2026). While this in-house, proprietary solution from a single company is undoubtedly easier to implement than a universal framework for published compositions and sound recordings across diverse platforms, publishers, and record labels, it still serves as a functional proof of concept. It represents a constructive step toward the healthy integration of artificial intelligence and the copyrighted materials used to inform its outputs, demonstrating a viable model where original creators are financially rewarded while technology providers maintain the utility of their tools.

Concluding Remarks

While proprietary, closed-loop systems like Splice offer a functional proof of concept for Training Data Attribution, they remain isolated solutions. For the broader music industry to endure the paradigm shift of generative AI, these micro-level attribution mechanics should be scaled up into macro-level statutory and licensing frameworks. As the industry confronts this profound disruption, the debate transcends technical definitions of copyright infringement to ask a more fundamental question about the value of human creation. While proponents of the "fair training" defense attempt to humanize algorithmic ingestion, equating a machine's data processing to a human reading a book or a DJ creating a musical pastiche (Torrance & Tomlinson, 2023) fundamentally ignores the commercial reality of artificial intelligence. Generative AI models are not autonomous, culturally participating individuals; they are inherently commercial products developed for profit by a select few corporate entities. The question is not merely whether unauthorized training can be semantically justified as "fair," but rather if we as society deliberately and consciously choose to structure our legal frameworks to disproportionately serve the proprietors of these machines at the expense of human creators. The purpose of copyright is not only to protect creators’ financial interests, but more importantly to incentivize artistic creation and prevent the irreversible depreciation of our foundational cultural capital, and treating corporate-owned algorithms as the legal equivalents of human learners could set a dangerous precedent. Instead, the developers of these models should be required to pay licensing fees to sustain the human intellect and property that taught the AI its capabilities. By implementing scalable, technological solutions such as Training Data Attribution (TDA) and user-centric blanket licensing, the industry could ensure that original creators are fairly remunerated, fostering an ecosystem where algorithmic innovation and human artistry can sustainably coexist.

2026 by the authors. This article is an open access article distributed under the terms and conditions of

the Creative Commons Attribution (CC BY) license .