Hey Soulless Sister: Regulating AI Music Under International Copyright Law

[Lara Manbeck is an Associate Legal Officer at the International Court of Justice. She holds a JD from Columbia Law School, an MPhil from the University of Cambridge, and a BA from New York University.

Abhijeet Shrivastava is a public international lawyer and holds an LLM from the University of Cambridge and a BA, LLB (Hons) from OP Jindal Global University.]

Introduction

The popular imagination has long contemplated a world in which the boundary between technology and humanity would become indistinct. Despite such foresight, artistic production has continued to be seen by many as an exclusively human, if not also a spiritual, endeavour (see, for example, Pope Leo XIV’s recent encyclical Magnifica Humanitas). In an era of increasingly undetectable deepfakes, however, it should be no surprise that this assumption is now in question. AI has now, perhaps irreversibly, intruded upon one of the most primordial realms of human expression: music.

Cursory scrolling through social media or music platforms readily demonstrates the prevalence of this phenomenon. The silence of long-deceased artists is being broken by generative AI that can replicate their voices, sometimes prompting legal action from their estates. AI-generated songs have permeated streaming services, with one study estimating that revenue from such music will increase from $100 million in 2023 to around $4 billion in 2028. Aptly concerned about the existential threat these developments pose, artists have assembled in protest against perceived regulatory lacunae.

The question thus arises as to whether technology has outpaced international regulation in the realm of music. Offering an initial response, this post identifies the central challenges posed by this possible gap. After providing an overview of the applicable law, we turn to issues of ambiguity and then explore whether existing exceptions can accommodate generative AI music, before finally assessing what reforms may be necessary going forward.

Core International Framework

Million-dollar lawsuits are ongoing in domestic courts against producers of AI-generated music, typically with copyright violations as the cause of action. However, as of writing, domestic regulators have yet to pointedly tackle the source of the problem. At the international level, there is likewise no dedicated international treaty regulating AI-generated music. Rather, as with earlier technological developments such as malicious cyber activities, the situation calls for the application of existing international legal norms to novel circumstances. 

In the realm of international intellectual property, this presently comprises a collection of long-standing copyright and ancillary treaties, including principally the Berne Convention for the Protection of Literary and Artistic Works, 1979 (“Berne Convention”); the Agreement on Trade-Related Aspects of Intellectual Property, 1994 (“TRIPS”); the Copyright Treaty, 1996 (“WCT”) under the World Intellectual Property Organization (“WIPO”); and finally, the WIPO Performances and Phonograms Treaty, 1996 (“WPPT”). 

The Berne Convention, which has 182 States Parties, sits at the foundation of the international intellectual property regime. Covering “every production in the literary, scientific and artistic domain, whatever the mode or form of its expression” (Article 2), it requires States Parties to provide certain minimum rights to authors. These include, among others, rights of reproduction (Article 9), public performance (Article 11), broadcasting (Article 11bis), and adaptation (Article 12). Moreover, Article 6bis recognizes “moral rights,” i.e., the right of the author “to claim authorship of the work and to object to any distortion, mutilation or other modification of, or other derogatory action in relation to, the said work, which would be prejudicial to his honor or reputation.”

Two basic features of the Berne Convention are of particular relevance to recent innovations like AI-generated music. For one, its safeguards apply automatically: a composer need not register their work abroad in order to benefit from treaty-based protection (p. 5). Secondly, copyright enforcement remains territorial (p. 85). In other words, while the Berne Convention imposes cross-border obligations on States Parties, infringement claims are generally resolved under the law of the country where protection is sought (p. 34). This means that the same act of AI training, or the same AI-generated output, may be treated differently depending on how a particular national jurisdiction addresses it, giving rise to a risk of fragmentation if States Parties cannot develop common ground rules. 

Further, TRIPS, the WCT, and the WPPT supplement the Berne Convention by incorporating its standards into the World Trade Organization framework, and by strengthening enforcement obligations (see TRIPS Articles 9(1), 41(1)); by clarifying authors’ rights in the digital environment, including the rights of communication to the public and making available (see WCT Articles 6, 8), and performers’ rights in their performances, including reproduction and making available (WPPT Articles 7, 10); and by protecting the related rights of phonogram producers (WPPT Articles 11-14).

Ambiguities Surrounding AI-generated Music

The international intellectual-property treaty regime protects distinct rights that different right-holders may hold in a given musical work. The same composition may be protected as both a musician’s authorial work and a basis for performers’ rights. Furthermore, the producer of the musical recording too might possess rights in the phonogram itself. International intellectual property law, therefore, provides multilayered protections to a range of stakeholders in works that may be appropriated through generative AI. 

Yet, this treaty regime developed long before the rise of AI, which presents problems that its drafters and ratifiers could not have had in mind. The legality of AI-generated music under this regime, therefore, raises novel questions. Answering them requires the consideration of multiple, interrelated issues, including the nuances surrounding the various stages of an artist’s creative process, as well as the nature of and extent to which their work is used in AI-driven music generation. 

The first point of analysis is whether, and when, generative AI models may infringe upon copyright to begin with. After all, it remains to be seen whether international law will treat AI training as analogous to the way in which artists draw inspiration from existing music, or instead as plain and simple copyright infringement. One might particularly ask whether the unfettered use of copyrighted music in training AI may infringe on the songwriter’s copyright, the performer’s interpretation of the song, and the rights of a producer over the recording. On first principles, the concern is straightforward: generative AI training usually involves the copying of protected works into datasets, their processing, and storage in digital form. Since Article 9 of the Berne Convention and national legislation across the world afford authors the exclusive right to authorize reproduction “in any manner or form,” training AI with protected music plausibly violates such a right. These “input” infringements are the first problem on the AI-music production line.

A second possible infringement arises at the “output” stage. By nature and design, AI models produce outputs that resemble the models’ training inputs. If a model outputs songs that resemble a copyrighted composition, a distinct though related infringement problem results: the output sounds like plagiarism. When a model generates a song that substantially reproduces protected expression from existing compositions or sound recordings, ordinary infringement law could offer appropriate remedies. Yet, the hard questions would be evidentiary and doctrinal. How close would be too close? Does the output copy protected expression, or merely imitate genre, style, mood, instrumentation, vocal quality, or production technique? How could a claimant prove dependence on a particular work where training data is opaque? Should there, thus, be an obligation to disclose training data? These difficulties are acute in the music industry, where similarity may result from commonplace chord progressions, rhythmic patterns, or genre conventions. 

As mentioned above, litigation is ongoing between record companies and generative AI companies over allegations of mass unauthorized copying of sound recordings. It is, therefore, too early to predict the direction in which the law will develop. Moreover, beyond the norms established under the international intellectual property regime, some national jurisdictions recognise additional protections, such as personality rights for celebrities with distinctly recognisable voices. These national legal doctrines will probably play an important role, particularly in disputes involving leading artists. In this vein, the interpretation and evolution of the international intellectual property framework is likely to be shaped by the outcomes of domestic legal developments and the legal principles that emerge from them. However, reliance on national judicial developments as the primary driver of international norms may risk producing a piecemeal, and potentially fragmented, system of governance. 

An Additional Grey Area: Lawful Exceptions

Thus far, we have provided an overview of the different stages at which, and different manners in which, generative AI music training and production could give rise to infringements of copyright protections provided under international intellectual property law. However, a critical question remains: might such reproduction fall within the scope of lawful exceptions? 

International copyright law does not create a general privilege for text and data mining, machine learning, or AI training. Instead, it sets the outer limits within which national exceptions must operate, most importantly through the Berne Convention “three-step test”: exceptions must be limited to certain special cases, must not conflict with the normal exploitation of the work, and must not unreasonably prejudice the legitimate interests of the right holder. 

Whether lawful exceptions protecting music-generative AI training or use pass the “three-step test”, therefore, depends on their contours. A narrow exception for non-commercial research would be easier to justify than a broad exception allowing commercial AI developers to ingest entire music catalogues without consent or remuneration. The issue becomes especially contested insofar as licensing markets for training data already exist or are emerging, and where trained systems may displace tthe very musicians, performers, and producers whose works made them possible. International law does not resolve this issue conclusively, leaving States considerable room to diverge.

Looking Ahead

The intentionally limited scope of this piece ought not to understate the scale and significance of the issue it examines. AI-generated music carries tremendous social, moral, and philosophical implications. UNESCO has forewarned the rise of a “monoculture by algorithm…as AI systems become increasingly embedded in cultural production, distribution and consumption.” The Human Rights Council’s Special Rapporteur on cultural rights has similarly cautioned that it is “necessary to ensure that artificial intelligence does not stifle human creativity.” Evidently, generative AI music raises profound concerns that warrant sustained scholarly and regulatory engagement beyond international copyright law.  

To address the challenges posed by problematic generative AI music, the most realistic path forward is likely not the creation of a single, comprehensive AI-music legal regime, though regulation at the international level is critical. While a broader treaty governing generative AI as a whole may eventually prove valuable, a more practical approach would involve a series of incremental reforms that can adapt to rapidly evolving technological developments. In concert with domestic regulators, the WIPO could progressively develop soft-law standards, and eventually treaty norms, addressing dataset transparency, rights-preservation protocols, licensing practices, rights-management information, and performer protection. States could pair any text-and-data-mining exception with lawful-access requirements, effective opt-out mechanisms, and collective or extended collective licensing where individual licensing is impracticable. Courts and legislatures could also develop burden-shifting rules: where a claimant makes a credible showing that its catalogue was ingested and that a model can produce close matches, providers should be required to disclose relevant training and filtering evidence.

All that said, society is changing rapidly. Some renowned artists worry that good musicking may soon be defined by one’s ability to devise the right kinds of prompts for generative AI. Perhaps the industry will eventually come to appreciate and encourage the lawful production of AI music. There are indeed discussions surrounding whether, when, and how, AI-generated music may itself enjoy copyright protections. Ultimately, however, copyright has always substantially rested on the idea that human creativity deserves recognition and protection; an assumption that is now being challenged as AI not just extends human creativity but potentially competes with it. If the mushrooming of AI music forces lawyers to recalibrate that premise, they must ensure that the pursuit of technological progress does not allow AI to obscure or exploit human artistic creation.