Music Law and Artificial Intelligence: From Cloned Artists to AI-Generated Works
The rise of artificial intelligence (AI) in the music industry is unleashing a revolution that is profoundly changing the way music is created. This development raises complex legal questions about AI and intellectual property law, including copyright (AI music copyright) and neighboring rights. How can we protect the creative rights of artists and composers while simultaneously allowing room for technological innovation?
Mauritz Kop, Partner at AIRecht
In this contribution, Mauritz Kop offers an accessible legal overview of the key issues surrounding AI and music in The Netherlands and beyond. We will cover, among other things, whether AI may train on protected material without permission, how different rights organizations (Buma/Stemra and Sena) are dealing with AI, the status of AI-generated musical works, the threshold of human creativity, protection against AI voice cloning via privacy laws and moral rights, contractual implications, new obligations under the EU AI Act, differences between European and American law, and current legal cases. This article is written in a mix of legal jargon and clear explanations, aimed at readers such as artists, composers, music publishers, labels, voice actors, producers, and AI companies who want to understand their legal standing.
AI Training on Protected Music and Video Material: Legal Frameworks and Debate
Can an AI model in the Netherlands and the EU simply train on copyrighted material (such as music or video) without the permission of the rights holders? This is the core question of a major ongoing debate.
European TDM Exceptions: In 2019, the EU introduced several exceptions for text and data mining (TDM) with the DSM Directive. These provisions—relevant for AI training—allow the use of protected material without permission under strict conditions:
● Article 3 (Scientific Research): Research and cultural heritage institutions may apply TDM to lawfully accessible material for scientific research purposes without prior consent.
● Article 4 (General TDM Exception): For commercial purposes, TDM is permitted on lawfully accessible content, but rights holders have an "opt-out" right. They can explicitly reserve their rights, for example, through machine-readable metadata in a file or in the terms and conditions of a website. If they do so, the TDM exception does not apply, and using the material for training constitutes copyright infringement. In other words, an AI company will almost always need to arrange for explicit licenses to train on commercially released music.
The Role of Buma/Stemra: The Dutch collecting society Buma/Stemra, which represents composers and lyricists, has taken a clear stance. In line with Article 4, they have declared a general opt-out for their entire repertoire. This means that no AI model may be trained on music affiliated with Buma/Stemra without a specific license. Buma/Stemra emphasizes that this ensures fair compensation for composers and lyricists.
Sena's Position: Sena, the organization for the neighboring rights of performers and producers in the Netherlands, has also adopted a cautious stance. They advise their rights holders to explicitly state in contracts that their recordings may not be used for AI training without permission. This proactive advice aligns with Buma/Stemra's opt-out: both copyright and neighboring rights holders are encouraged to actively prevent their work from being used in AI without authorization.
Music Law and Artificial Intelligence: From Cloned Artists to AI-Generated Works
EU AI Act: Transparency Obligations and System Monitoring
In July 2024, the European Union adopted the AI Act, a comprehensive regulatory framework that is also relevant to music AI. While this law does not explicitly modify copyright permission requirements, it does impose new duties on AI providers, particularly concerning transparency. For example, the EU AI Act (effective from August 2025) obliges developers of generative AI models to:
1. Disclose Training Data: Provide a sufficiently detailed summary of the copyrighted works used for training the model. This is a significant obligation because it forces AI companies to be transparent about their data sources. Rights holders can use this information to check whether their material has been used and, if so, whether this was done lawfully (i.e., with a license or under an exception).
2. Label AI Content: Clearly indicate that content (such as music, images, or text) has been generated or manipulated by an AI system. This helps to prevent deception and ensures that users know they are interacting with AI-generated output.
3. Implement a Copyright Policy: Establish and maintain a policy to respect EU copyright law. This includes demonstrating that they have made efforts to obtain licenses for the data they use, especially when the TDM opt-out has been exercised by rights holders.
Although the AI Act does not create a new "training right," its transparency requirements give rights holders a powerful tool for enforcement. It will become much easier to identify and act against AI systems that have been trained on protected music without permission.
US Law: The "Fair Use" Doctrine
The legal situation in the United States is fundamentally different due to the fair use doctrine. This doctrine allows for the limited use of copyrighted material without permission for purposes such as criticism, commentary, news reporting, teaching, or research. AI companies often argue that training their models on large datasets constitutes "transformative" fair use because the goal is not to reproduce the original works but to learn patterns from them. Whether this argument will hold up in court is the big question in several ongoing US cases. In Europe, the scope for such an appeal to fair use is minimal; the rules are stricter and more explicit, giving rights holders a stronger position here (through the opt-out and the AI Act's transparency obligation) than in the US.
Composition, Lyrics, and Master: Different Music Rights Regimes
In the Netherlands, music rights are broadly divided into copyrights (for the composition and lyrics of a song) and neighboring rights (for the performance and recording, i.e., the master/phonogram). AI applications can infringe on all of these rights, so it is important to examine them separately—as well as the roles of organizations like Buma/Stemra (for copyright) and Sena (for neighboring rights).
Composition and Song Lyrics (Copyright)
The melody, harmony, and rhythm of a musical work—in short, the composition—and also the lyrics fall under traditional copyright. The composer and lyricist are the copyright holders of their own original creations. A fundamental principle is that a work is only protected by copyright if it is a human's original creative result, bearing the personal stamp of the creator. This follows from established EU case law (e.g., the Infopaq ruling) and is enshrined in the Dutch Copyright Act.
Buma/Stemra's stance is firm: the use of their copyrighted repertoire for AI training without prior permission is not allowed. They have even established registration procedures for (partially) AI-generated musical works, acknowledging that hybrid works (human + AI) will occur but that they must be properly declared. The message to AI companies is clear: without a license, there is nothing to be gained from our catalog.
Master Recordings and Phonograms (Neighboring Rights)
In addition to the copyright on the composition and lyrics, there are the neighboring rights for performers and music producers. As soon as a song is recorded, a right is created for the performer (e.g., the singer/band) and for the producer (the label) on that phonogram (sound recording). These neighbouring rights last for a shorter period than copyrights (usually 70 years after the recording) but are crucial for exploitation: they ensure that artists and labels get paid when their music is played or sold.
AI and Neighboring Rights: An interesting question is how neighboring rights play out when music is created with the help of AI. Suppose an AI system generates a new song "in the style of" a certain artist and produces a recording of it (for example, through a voice clone of that artist, see below). Who is then the "performing artist" with rights to that recording? The traditional definition—a human delivering a performance—falls short here. It is possible that such an AI recording would have no neighboring rights, unless the law is amended to create a form of right for such scenarios. This topic is still in its infancy but touches on the core of what a "performance" is in a legal sense.
Sena, which manages neighboring rights, has stated that anyone who plays AI-generated music in public (e.g., in a café or store) must have a Sena license. Their reasoning is pragmatic: if a recording is commercially exploited and made public, it should be treated as a regular phonogram for which fees must be paid. Sena also advises its rights holders to make an explicit reservation against the use of their repertoire for AI training. This preventive advice aligns with Buma/Stemra's opt-out: both copyright and neighboring rights holders are advised to actively prevent their work from ending up in AI without consent.
Differences between Buma/Stemra and Sena Policy on AI-Generated Musical Works
Buma/Stemra and Sena apply different principles and roles when dealing with AI-generated music, stemming from their legal mandates and the nature of the rights they manage.
Buma/Stemra: Focus on Copyright (Composers and Lyricists)
○ Represents the interests of composers and lyricists. They collect and distribute royalties for the use of the copyrighted work (the composition and lyrics).
○ For AI-generated musical works, Buma/Stemra's emphasis is on whether there is human creativity. Only when a human has sufficient creative input can a work fall under copyright and thus under the management of Buma/Stemra.
○ Buma/Stemra states that explicit permission is required for the use of their repertoire for AI training.
○ For works generated entirely by AI without human creative contribution, Buma/Stemra generally sees no copyright protection, and therefore no role for itself as a collective management organization.
Sena: Focus on Neighboring Rights (Performing Artists and Producers)
○ Manages the neighboring rights of performing artists and producers of phonograms (sound recordings).
○ Sena continues to collect for phonograms released for commercial purposes, even if they were created with the help of AI. The method of creation (human or AI) is less relevant: if a recording is available as a phonogram, it is treated as such, and collection occurs upon public performance.
○ Sena advises rights holders to make an explicit reservation against the use of their repertoire for AI training but does not (yet) have a policy that excludes AI-generated phonograms from collection and distribution.
○ Sena focuses primarily on exploitation: as soon as a track (even AI-generated) is commercially released and available, it is included in the collection and distribution of royalties.
Differences between Buma/Stemra and Sena Policy on AI-Generated Musical Works Payments
Conclusion: Buma/Stemra requires explicit permission for AI training on their repertoire and only recognizes works with human creative input, while Sena focuses on the exploitation of recordings and treats AI-generated phonograms as regular sound recordings as long as they are commercially available. The core difference lies in the nature of the rights and the role of human creativity versus exploitation.
Copyright Status of AI-Generated Music
A crucial legal question is whether a musical work created by an AI system can be protected by copyright at all. According to European law (and US law), copyright protection requires a work to be the result of human creative choices. A work that is generated entirely by an AI, without decisive creative intervention from a human, is generally considered to be in the public domain. This means anyone can use, copy, and distribute it freely.
● Hybrid Works: In practice, many cases will not be black and white. Often, a human musician will create a song together with AI—for example, someone who directs an AI tool and creatively processes its output. In such cases, the work may be copyrightable, provided the human's contribution is sufficiently original and creative. The question then becomes how to determine the threshold for this "creative contribution." It must be more than simply pressing a button; the human must have made free and creative choices that are reflected in the final result.
● Neighboring Rights & AI Performances: A separate point of concern is whether neighboring rights can arise from AI performances. Suppose an AI system sings or plays a new song (whether through voice cloning or not). Who is the performing artist with a right to that performance? Traditionally, one would say: no one, because there is no natural person performing. The consequence could be that AI music recordings have no performer's rights—a potential gap in income for the music industry. A solution could be to still consider the role of the human who operates the AI system or provided the data input as the performer. However, this is legally uncharted territory. Sena's current view seems to be implicitly that, as long as the result is brought to market as a phonogram, they will treat it as something for which collection takes place for the benefit of something or someone. But strictly legally, a redefinition may be necessary if AI performances become commonplace.
Voice Cloning by AI: Personality Rights and Privacy
The technology of AI voice cloning (synthetically imitating someone's singing or speaking voice) has taken off. AI systems can now generate lifelike imitations of specific artists or voice actors, often without the public hearing the difference. This raises not only copyright issues but also touches on the personality rights and privacy of performers. After all, a voice is something unique and personal—its unauthorized use or imitation can feel like an infringement on one's identity.
● Legal Protection Against Voice Misuse: What options do artists and voice actors have to defend themselves against this? In the Netherlands (and the EU), they can invoke the right to privacy/private life (Article 8 ECHR, elaborated in Article 10 of the Dutch Constitution) and personality rights. A person's voice falls within their personal sphere; using it without permission can be seen as an unlawful act against the artist. In addition, trademark law (if the artist has registered their name/voice as a trademark) and the right of publicity (portrait right) by analogy come into play, although a voice is not a "portrait," the issue concerns the commercial use of one's identity.
In practice, however, it is legally challenging. There is (as yet) no explicit "voice right" law. Artists must therefore seek redress through indirect routes, for example, by arguing that their voice is a personality right that is being violated or that the public is being misled (an unfair commercial practice). And although alarming examples of misuse (e.g., deepfake pornography or fake endorsements) lead to calls for stricter regulation, legislation in this area is still in development. In the US, for example, Tennessee's ELVIS Act has been introduced, which explicitly protects artists' voices from unauthorized AI imitation. It is likely that similar laws will follow in Europe.
Music Contracts in the AI Era
The rise of AI also forces a revision of music contracts and licensing agreements. Existing contracts between artists, labels, publishers, etc., are generally not written with AI in mind. As a result, they are silent on questions such as: Can a label allow an artist's catalog to be used by an AI for training? Who owns the rights to a track that is partly composed by AI? It is crucial to fill these gaps with new clauses.
● Licenses for AI Training: An increasing number of artists and composers want to explicitly state in their contracts whether or not their work may be used for AI purposes, via an AI training license. Some opt for a hard prohibition clause: their label or publisher may not put their music into a dataset without permission. Others also see opportunities and are negotiating for remuneration for AI use—for example, a license agreement that AI companies must enter into to be allowed to train on the repertoire. It is clear that permission and compensation for AI use is becoming a new area of negotiation in the music sector, either individually or collectively.
● AI Output and Authorship: Another contractual point is the ownership of AI-generated output. Suppose a producer uses an AI tool to create a beat or melody line that ends up in a new song. Who does that contribution belong to? Contractually, you can agree that all output from the AI is still considered the work of the human creator (as far as possible), or that the rights accrue to the party that deployed the AI. In songwriting camps, experiments are already being conducted with agreements that if AI tools are used, all participants share pro-rata in any rights or income. These agreements must be tailor-made, as the law does not yet provide a fixed framework here.
Europe vs. America: Different Legal AI Copyright Tracks
In the field of AI and copyright, we see clearly divergent approaches in Europe versus the United States. Europe is trying to find a balance between innovation and protection through regulation (such as the AI Act) and adjustments to copyright law. The US, on the other hand, relies on existing doctrines and leaves it primarily to the courts and the market to set boundaries.
Some differences at a glance:
● Authorship of AI Works: As mentioned, the US currently holds the line that works generated entirely by AI do not get copyright—period. Europe has not explicitly established this, e.g. by clarifying the EU Copyright Act, but implicitly follows a similar course (human creativity required). At the same time, discussions are taking place in EU circles about possible new rights or other solutions, whereas the US is less inclined to do so.
● Fair Use vs. Licensing: In the US, the unauthorized use of protected content for AI training may be justified under fair use (provided it is transformative, limited, not market-disrupting, etc.). In Europe, there is no fair use doctrine; here you must either fall within a strict exception or have permission. The EU opt-out is stricter than the American practice, where data is often used tacitly unless someone complains.
Ongoing Lawsuits and Enforcement
In both Europe and the US, rights holders and regulators are beginning to go to court to set limits on unauthorized AI use. Some prominent ongoing cases in the field of AI and music include:
● Meta (Facebook/Instagram) – AI training on user data (EU): In 2025, an attempt was made in Germany to stop Meta from using the Facebook/Instagram posts of European users for AI training without explicit consent. A consumer rights organization (Verbraucherzentrale NRW) sought an injunction, but the Cologne Higher Regional Court refused this in May 2025. Meta is for now allowed to continue training, provided users have an opt-out option. At the same time, a privacy investigation is underway: the Hamburg data protection authority is examining whether Meta's approach complies with the GDPR.
● Music Publishers vs. Anthropic (US): In 2023, several major music publishers (including Universal Music, Concord, and ABKCO) sued the AI company Anthropic. They allege that Anthropic's AI model, Claude, was trained on their copyrighted song lyrics without a license and can reproduce them almost verbatim upon request. In March 2025, a judge ruled that the publishers had provided sufficient evidence for the case to proceed, rejecting Anthropic's motion to dismiss. While Anthropic won an early round by arguing that the publishers had not sufficiently specified which songs had been infringed, the publishers "are very confident in their case" and are proceeding with the main lawsuit. The outcome of this and similar cases (there are also author and visual artist cases against other AI companies) will be trend-setting. The American judgment on when AI training constitutes infringement will also influence how such issues are assessed in Europe.
Conclusion: A Balance Between Innovation and Rights – Be Prepared
The intersection of AI and music law is a dynamic and complex field where technology is advancing faster than legislation can keep up. Nevertheless, the contours of a new legal framework are becoming visible. In Europe, the emphasis is on permission, transparency, and compensation. Rights holders are given tools (the TDM opt-out, the AI Act) to maintain control over their work, while AI developers are pushed towards a licensing-based model.
For music professionals—from independent artists to major labels—it is essential to keep a finger on the pulse. Rights in AI-generated music must be properly secured by contract. Artists would be wise to explicitly state whether their voice or tracks may be used for AI training. Publishers and labels need to update their agreements to create clarity about ownership and liability concerning AI creations. And AI companies in the music sector will have to work with transparency and licenses if they want to avoid conflicts.
These developments are happening at a rapid pace. New EU rules (like the AI Act) are just around the corner, and legal precedents are being set now. It is therefore not an unnecessary luxury to seek legal advice on AI and music from specialists. An experienced music law attorney can help to map out both opportunities and risks—from protecting your voice and songs against misuse to optimally licensing AI technology. In this way, as a creator or enterprise, you can benefit from AI innovations with peace of mind, within the boundaries of the law. In the AI era, more than ever, the saying applies: look before you leap—but with the right knowledge and legal support, creative control remains in your hands.