Music law is on the front lines when it comes to generative AI. This is one of the key areas of focus for the International Association of Entertainment Lawyers (IAEL, www.iael.org), which has been publishing edited books for forty years—to which we have contributed. In particular, we co-edited and co-authored The AI Revolution: The International Legal Perspective (IAEL, 2025) and contributed to The International Music Business Navigator for Lawyers and Artists (IAEL, 2026) with a practical guide to artist contracts presented at the Amsterdam Dance Event (ADE).

Current events give these works immediate relevance: on November 3, 2026, Advocate General Maciej Szpunar presented his opinion in the case of Like Company v. Google Ireland (C-250/25), the first case brought before the CJEU concerning generative AI and the law. While the dispute specifically concerns the neighboring rights of press publishers, the opinion also addresses issues of interest to all IP rights holders, particularly regarding the use of their protected content to train AI models and its reuse by generative AI systems.

At the same time, European and French initiatives are multiplying to regulate the use of works in model training. Meanwhile, another question remains: what legal status should be assigned to generated content?

I. Model Training: The Like Company v. Google Case Before the CJEU and Legislators’ Responses
1. The Reference for a Preliminary Ruling
The first question concerns training: does copying works to train a model constitute reproduction, and can the text and data mining exception apply?

The CJEU is hearing this matter in the case of Like Company v. Google Ireland (C-250/25), referred by a Hungarian court as part of the first case involving generative AI and copyright brought before the Court . Like Company, which operates Hungarian news portals, alleges that Google used its articles through its Gemini chatbot without authorization. It does not invoke copyright but rather the neighboring rights of press publishers.

The main provisions at issue are: Article 15 of Directive (EU) 2019/790, concerning the neighboring rights of press publishers, Articles 2 and 3 of Directive 2001/29/EC on reproduction and making available to the public, and Article 4 of Directive (EU) 2019/790 establishing the text and data mining (TDM) exception, subject, in particular, to rights expressly reserved by rights holders in an appropriate manner.

For the music industry, the issue at hand directly concerns the conditions under which musical works, lyrics, and recordings may be used to train models without the authorization of rights holders.

The referral concerns both the training of the chatbot—to determine whether the analysis of content constitutes a reproduction that may fall under the TDM exception—and its responses, to assess whether the reproduction of identical passages constitutes a reproduction or a making available attributable to the provider. The upstream and downstream aspects are thus combined in a single case.

Like Company argues that the use of its content is general, continuous, and systemic, and that publishers therefore cannot be required to provide proof of each individual interaction. Google responds that the similarities are accidental or the result of hallucinations, invokes the exceptions for temporary acts of reproduction and text and data mining, and argues that Gemini produces new content rather than copies.

2. The Advocate General’s Opinion
The highly anticipated opinion of Advocate General Maciej Szpunar, delivered on November 3, 2026, distinguishes between upstream and downstream processes. He rejects the Commission’s argument for blanket inadmissibility , but deems the questions regarding training to be hypothetical.

On the one hand, there is no evidence that Like Company’s articles were used to train Gemini: the disputed summary could instead be explained by real-time retrieval of the article, as the model was not deployed until December 2023. On the other hand, training does not in any case constitute “online use” within the meaning of Article 15 .

For the sake of completeness, the Advocate General nevertheless provides, on a provisional basis, guidance that may be of interest to all rights holders :
• Training involves reproductions.
• The exception under Article 4(3) of Directive (EU) 2019/790 applies in principle, provided that access is lawful, there is no reservation of rights expressed in machine-readable form, and measures against reposting are in place (see Part II).
• He is hesitant to extend EU law to training conducted outside its territory, such as here in the United States, and calls on the legislature to decide the matter.
• It views the transparency requirement in Article 53 of the AI Act as a necessary complement to the exception and acknowledges that Member States may provide compensation for creators .

The issue of copyright notices has also been brought before the Court in another case, brought by a Hungarian publisher (Mobilguru v. Google, C-829/26), specifically regarding the requirements applicable to machine-readable notices .

3. Responses from European and French Legislators
These guidelines are part of a framework that lawmakers are developing in parallel. At the European level, Article 53(1)(c) and (d) of AI Act (EU) 2024/1689 requires providers of general-purpose AI models to comply with Union copyright law, including the copyright notices set forth in Article 4(3), and to publish a summary of the training data.

The European Parliament goes further in its resolution of March 10, 2026 (Axel Voss report). It asserts that EU copyright applies to models marketed within its territory, regardless of where the training took place; proposes a registry of works used, hosted by the EUIPO; and calls on the Commission to examine the appropriateness of equitable remuneration for past- uses. This resolution has no legislative effect, but the Commission welcomed it, noting the upcoming evaluation of Directive (EU) 2019/790. For music industry stakeholders, such a registry could, in particular, facilitate the identification of works and recordings used to train models.

In France, the difficulty of proving that a work was used to train a model is at the heart of Senator Laure Darcos’s proposal, which consists of a single article . It establishes a simple presumption of use whenever there is evidence suggesting that a work has been exploited. Such evidence may relate to the development or deployment of the system or be present in the generated output. The provider may rebut this presumption by demonstrating in court that the content was not used. The French Council of State (Conseil d’État) has clarified that this is a presumption of use and not of infringement: it is then up to the judge to determine whether infringement has occurred .

Unanimously adopted by the Senate, the proposal is expected to be considered in a session of the National Assembly during the week of November 19 .

These mechanisms are particularly important for owners of music catalogs, who face difficulties in proving that their work has been reproduced by AI providers. They could thus facilitate the establishment of proof without exempting the claimant from subsequently demonstrating the existence of an infringement of their rights.

However, there remains the scenario where the output does not reproduce the work but rather “synthesizes,” “recombines,” or “aggregates” it into new content: terms that AI providers use to assert the absence of reproduction. The rights holder must then still demonstrate that what has been incorporated falls within the scope of the protected form and not within the realm of ideas, facts, or styles, which remain in the public domain. It is in this gray area that the difficulty of proving infringement is most acute.

Another development could take place at the European level. According to a working draft of the annexes to the European Commission’s 2027 work program, published by Contexte prior to its presentation on November 4, 2026, the Commission is considering a “digital transformation, simplification, and copyright package,” expected in the second quarter of 2027. The use of the omnibus approach could allow for a review of Directive (EU) 2019/790 and a reexamination of the rules governing the use of works protected by AI systems. This prospect is of direct interest to music catalog owners, whose licensing and remuneration terms could be affected.

II. The Output of Works: Storage, Reproduction, and Provider Liability
There remains the issue of the model’s output: when it produces an AI-generated work, to whom is the act attributable?

“Storing” refers to a model retaining specific passages from its training data within its parameters. It can then reproduce them almost identically in response to an appropriate user query. In the music industry, this issue concerns lyrics as well as melodies and recordings.

In late 2025, the Munich Higher Regional Court (OLG München) ruled against OpenAI for the reproduction of song lyrics. The court found that both memorization within the models and reproduction in the output were established, refused to attribute liability to users, and did not consider the exception for text and data mining to be a sufficient defense. It acknowledges in principle the distinction between learned parameters and stored copies but finds, in practice, that the models reproduced the lyrics almost identically .

The same court extended this ruling in July 2026, in the first European decision concerning a music generator. It largely granted the requests for an injunctive relief, disclosure of information, and damages filed by GEMA against Suno regarding six known works, as the lyrics were not at issue. The court noted that these works had been stored in versions v3.5 and v4 of the models and that liability rested with Suno, not its users .

These decisions, which are not binding on the CJEU, illustrate the risks specific to music generators: the recognizable reproduction of protected lyrics or melodies may constitute a reproduction attributable to the provider.

In his opinion regarding the Like Company case, however, the Advocate General urges caution . In his view, the memorization results from overfitting—a byproduct of the training process that is not entirely avoidable—and the output is generated in response to a user’s request. The parameters, taken in isolation, are merely preparatory material: to regard them as a reproduction would subject the model itself to exclusive rights.

On the other hand, an output that is identical or nearly identical to a work is likely to constitute a reproduction, to be assessed on a case-by-case basis. The difference with Munich therefore concerns the location of the reproduction (in the model or in the output), with both approaches agreeing that the output is actionable.

Regarding the use of the chatbot—which he places at the heart of the case—the Advocate General proposes the following criteria :
• Reproduction. Neighboring rights protect the form of expression of the articles (their layout, structure, and wording), not the facts reported. Reproduction occurs if the chatbot has unlawfully accessed the publication—in particular by circumventing technical measures—or if this form is “recognizable” in the response to the extent that the user can dispense with consulting the original.
• Availability to the public. This is established only if such recognizability is systematic in responses to related queries.
• Attributability. The act is, in principle, attributable to the user who initiates it. It is also attributable to the provider when it is inherent to the provider’s business model or when the provider has failed to take the measures expected of a diligent operator, such as enforcing access restrictions or output filters. The Advocate General rules out the implicit consent of a publisher who leaves its pages accessible.

These criteria, formulated in the context of the neighboring rights of press publishers, nevertheless call for an examination of the recognizable reproduction of a protected work as well as the preventive measures put in place by providers.

III. The Status of AI-Generated Works: No Protection Without Human Creation
What is the status of content produced by AI: is it protected, and under what law? The CSPLA (High Council for Literary and Artistic Property) has dedicated a report to this topic, focusing on the “output”—that is, what generative AI systems produce—and referring the issue of “input” to other studies . It thus complements Case C-250/25 and the Darcos proposal, which concerns what is fed into the model.
The report distinguishes between two categories:
– Hybrid creation is one in which humans retain a creative role: it may be protected if the threshold of originality is met.
– Synthetic production is generated without significant human intervention and falls outside the scope of copyright.

The report rules out the creation of a specific copyright or a sui generis right for such works. It places them in the public domain upon their generation, without moral rights, due to the absence of an author. It considers, only prospectively, a paid public domain, conceived as a tool to steer the market in favor of human creation.

This distinction directly affects musical works: a piece composed by AI without creative human intervention would not be protected by copyright, unlike a creation in which humans retain control over free and creative choices.

The report bases this distinction on the CJEU’s Mio and Konektra judgment (December 2025) , according to which a work is defined as an object that reflects the free and creative choices of its author, as well as on the ruling of the First Civil Chamber of the French Court of Cassation (Cour de cassation) dated April 9, 2026 (Evema v. Someva) , which holds that the creative nature of choices cannot be presumed. In this regard, the report examines four stages (conception, creation, selection, and editing) to assess the creative role actually played by humans in music production.

Regarding the burden of proof, the report adopts a two-step process: the defendant presents evidence, and it is then up to the claimant to establish their creative choices, in accordance with German court decisions.

For music professionals, this distinction determines, in particular, whether they can claim copyright in a generated track or receive royalties as authors.
A contrast emerges between the two evidentiary mechanisms: the Darcos proposal establishes a presumption of use of the input works, whereas the CSPLA report rejects any presumption of originality regarding the output.

IV. From Generation to Distribution: Synthetic Tracks, Quotas, Neighboring Rights, and Transparency
It is at the distribution stage that these issues take on concrete significance. The CSPLA report quantifies this using secondhand figures: Spotify is said to have removed 75 million AI-generated tracks by the fall of 2025; 44% of new songs uploaded to Deezer are projected to be AI-generated by April 2026; and, according to an Ipsos study, 97% of listeners would not be able to distinguish an AI-generated track from a human creation .

1. Quotas for French-language songs
The debate over quotas is part of Senator Laurent Lafon’s proposal on the modernization of radio, which was adopted at first reading by the Senate and must now be reviewed by the National Assembly. According to the report from the Culture Committee, negotiations are underway between radio stations and the music industry regarding limits on song rotation, and the rapporteur hopes for a swift resolution so that these provisions can be incorporated into the proposal.

Added to this is the broadcast of AI-generated tracks. The music industry is asking that these not be played in very high rotation and is calling for transparency toward the public. ARCOM (Regulatory Authority for Audiovisual and Digital Communication) and the National Music Center addressed the subject in their consultation on musical diversity . However, the quota presupposes a work and an artist and draws on the concept of “new talent.” A purely synthetic track not only lacks an author but also circulates freely. This analysis sheds light on the call to limit its airplay, which aligns with the idea of a paid public domain, with the cost borne by whoever broadcasts the synthetic content.

2. Neighboring Rights and Equitable Remuneration
The issue of neighboring rights is more delicate. Article L. 213-1 of the French Intellectual Property Code (CPI) protects the producer of phonograms, who has the initiative and responsibility for the first fixation of sounds, regardless of any originality. According to the letter of the law, this right could therefore arise for a generated track.

The report, however, recommends excluding investment-based neighboring rights for purely synthetic productions: these rights compensate for an investment and an initial fixation that would not be achieved by merely writing instructions for an AI.

According to the report, phonogram producers themselves exclude any collection of royalties—whether for communication to the public or broadcasting—on recordings generated without substantial human intervention. It thus aims to prevent productions lacking substantial human intervention from capturing revenues intended to compensate phonogram producers for their investments.

It also notes that purely synthetic music falls outside the scope of the equitable remuneration provided for in Article L. 214-1 of the CPI. This remuneration is due upon the public performance and broadcast of phonograms published for commercial purposes, and its basis for calculation presupposes the existence of such a phonogram. Venues with sound systems are already refusing to pay on these grounds, raising concerns about a crowding-out effect to the detriment of performing artists and producers. However, the report targets these venues, not radio stations: the question of whether this reasoning applies to radio broadcasting remains open.

3. Streaming: The Example of “Fake Listening”
The risk of misappropriating collective remuneration also exists in streaming. At the request of the National Union of Phonographic Publishers (SNEP), the Paris Judicial Court (Tribunal judiciaire de Paris) ordered OVH and its hosting client in October 2025 to block access to two websites selling fake plays on Spotify, SoundCloud, and YouTube. The court found that these artificial streams automatically reduced the royalties of other artists in the “market-centric” distribution system—where royalties depend on a track’s share of streams on the platform—while unduly boosting the track’s ranking in charts and playlists .

Although this decision does not directly concern AI, it illustrates a similar risk: that of seeing artificially promoted content capture a disproportionate share of collective revenues at the expense of artists and producers whose works actually generate an audience. The massive influx of synthetic tracks makes this issue particularly sensitive for platform compensation models.

4. Transparency
The industry’s call for transparency is supported by Article 50 of the AI Act, which requires the artificial origin of content to be disclosed, and by the code of practice on the transparency of generated content. The report concludes that no single labeling technique is sufficient on its own. A broad coalition within the music industry has announced a voluntary labeling system for recordings that distinguishes between “AI-generated” and “AI-assisted” content. The market is also organizing on the supplier side: Suno has announced the use of audio watermarks and digital fingerprints, after entering into a licensing agreement with Warner Music . The report, however, tempers expectations regarding these measures: labeling certifies artificial origin, not copyrightability, and is “neither necessary nor sufficient.”

The challenge is therefore twofold: to enable professionals and the public to identify tracks generated or assisted by AI, while avoiding confusion between this information and the recognition of copyright or neighboring rights in the recording.

Conclusion
From classification to distribution, a consistent logic emerges: without transparency regarding what goes into the model and what comes out of it, neither rights holders nor the public can assert their interests. The Advocate General’s opinion has set the terms of the European debate. Model learning is subject to a framework whose territorial scope the Advocate General calls on the legislature to clarify, and the use of chatbots is assessed based on the recognizability of the form of expression and the provider’s due diligence.
For the music industry, the stakes thus go beyond the mere protection of works: they also concern the remuneration of authors, performers, and producers, the distribution of revenues from broadcasting, and the identification of synthetic productions.

The upcoming milestones will be decisive: the CJEU’s ruling in the Like Company case, the National Assembly’s review of the Darcos and Lafon proposals, as well as the European initiative on digital transformation, simplification, and copyright scheduled for 2027. These developments will help determine whether the existing legal framework is sufficient to protect the interests of authors, performers, publishers, and phonographic producers, or whether more far-reaching reform is needed.