Licensing, not litigation: the label settlements and the GEMA rulings

Licensing, not litigation: the label settlements and the GEMA rulings

By Dr. Arjun Sharma, Head of Neural Network Development, RadioPro LLC

Eighteen months ago the recorded music industry's answer to generative audio was a pair of complaints. On 24 June 2024 Universal, Sony and Warner, coordinated by the RIAA, sued Suno in Boston and Udio in New York, alleging "willful copyright infringement on an almost unimaginable scale" and seeking statutory damages of up to 150,000 dollars per work. (Suno complaint, D. Mass. 1:24-cv-11611) By the end of 2025 two of the three plaintiffs had signed licences with the companies they had sued. This article sets out what changed, what the German courts decided in the same months, and what the terms of the new agreements say about how a music model is now expected to behave.

The settlements and their terms

Universal and Udio, 29 October 2025. The first domino. UMG and Udio announced agreements that settled the litigation and licensed UMG's recorded music and publishing for a new subscription platform launching in 2026. The existing Udio service was to be kept running during the transition as a "walled garden," amended with "fingerprinting, filtering, and other measures." Udio stopped allowing users to download the songs they made, immediately. The new product was described as letting users customise, stream and share music "within what the companies describe as a 'licensed and protected environment.'" (AP, 30 October 2025) The next day UMG announced a separate alliance with Stability AI to build professional creation tools, with UMG's Michael Nash stating that "we will only consider advancing AI tools and products based on models that are trained responsibly." (Stability AI, 30 October 2025)

Warner and Udio, 19 November 2025. Warner's release describes a "landmark agreement that resolves the companies' copyright infringement litigation and establishes a clear framework for the development of Udio's licensed AI music creation service, set to launch in 2026." The service will be "powered by generative AI models trained on licensed and authorized music," and will let users make "remixes, covers, and new songs using the voices of artists and compositions of songwriters who choose to participate, while ensuring artists and songwriters are credited and paid." (Warner Music Group, 19 November 2025)

Warner and Suno, 25 November 2025. Six days later Warner settled with Suno on terms that go further into the product. New licensed models are to launch in 2026, and "When the new models launch in 2026, the current models will be deprecated." Songs made on the free tier "will not be downloadable and will instead be playable and shareable," and paid users "will have limited monthly download caps." Artists and songwriters "will have full control over whether and how their names, images, likenesses, voices" and compositions are used. Suno also bought Songkick from Warner. Robert Kyncl, Warner's chief executive, summarised the principle: "AI becomes pro-artist when it adheres to our principles: committing to licensed models" and opt-in for identity. (Warner Music Group release via PR Newswire, 25 November 2025)

The independents followed. On 20 January 2026 Merlin, the licensing body for independent labels, agreed that "Udio will develop new AI models trained on licensed music from participating Merlin members." (Merlin via IMPALA, 20 January 2026)

Sony Music has not settled with either company. It remains the plaintiff in both cases, and the Boston court has yet to reach the fair use question.

What the terms have in common

Read together, the agreements specify five conditions. They are now the market's definition of an acceptable music model.

  1. Licensed training data. Every agreement replaces the models that were sued with models trained on licensed and authorised material. Warner's Suno deal retires the old models outright.
  2. Opt-in for identity. Names, likenesses, voices and compositions are used only where the artist or songwriter chooses to participate.
  3. Credit and payment. Participation is tied to attribution and to a share of revenue.
  4. Controlled outputs. Downloads are removed or capped. Creations stay inside a licensed environment. Fingerprinting and filtering sit at the exit.
  5. Deprecation of the unlicensed past. The old systems are kept only for a transition and then switched off.

None of these conditions concerns the quality of the music. All of them concern provenance and control.

Munich: memorisation is reproduction

While the American plaintiffs were settling, a German court was ruling. On 11 November 2025 the Landgericht München I decided GEMA's case against OpenAI over nine German song lyrics, including "Atemlos" and "Über den Wolken." It was undisputed that the model had been trained on the lyrics and that the chatbot had reproduced them on simple prompts in 2024. The court largely granted GEMA's claims for an injunction, information and damages. (Landgericht München I, press release 11/2025)

The reasoning is the part that matters for anyone who builds models. The court held that storing the lyrics in the model's parameters is itself a reproduction under the InfoSoc Directive and section 16 of the German Copyright Act, and that encoding the work as probability values does not change that. The text and data mining exception covers the preparatory copying needed to build a training dataset, but not the reproduction of works inside the model. And OpenAI, not its users, is responsible for the outputs. As the legal daily LTO reported, the court ordered OpenAI "to refrain from storing the lyrics and outputting them in its models, to pay damages and to hand over information about the use and the revenue earned from it." (LTO, 11 November 2025) GEMA's chief executive Tobias Holzmüller said the internet "is not a self-service shop and human creative work is not a free template." (GEMA, 11 November 2025) The judgment is not final. OpenAI lodged an appeal on 8 December 2025.

GEMA's parallel case against Suno concerns six compositions, among them "Atemlos durch die Nacht," "Rasputin" and "Forever Young." The oral hearing, originally set for 26 January, was moved to 9 March 2026 after Suno filed a motion to disqualify the judges. (Landgericht München I, press release 1/2026) If the chamber applies the same reasoning to audio that it applied to text, a model that can be made to reproduce a recognisable composition is a model that contains a reproduction of it.

The American picture is narrower than it looks

Two decisions in June 2025 are often cited as a green light for training, and both are narrower than the headlines. In Bartz v. Anthropic, Judge Alsup found that training on lawfully acquired books was "exceedingly transformative" fair use, but that building a library from pirated copies was not, and the piracy claims were settled for 1.5 billion dollars. (ArentFox Schiff on the June 2025 order) In Kadrey v. Meta, Judge Chhabria granted Meta summary judgment while writing that the ruling "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful," and that a market dilution theory "will often cause plaintiffs to decisively win the fourth factor." (Loeb & Loeb on Kadrey v. Meta)

The US Copyright Office's report on training, released in pre-publication form in May 2025, drew the same line. "Various uses of copyrighted works in AI training are likely to be transformative," it wrote, but "making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries." Its policy recommendation was that "licensing markets should continue to develop." (US Copyright Office, Part 3, May 2025) The settlements of the autumn are that market developing.

On authorship, the D.C. Circuit held in March 2025 that "the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being," and Dr. Thaler's petition to the Supreme Court is pending. (Thaler v. Perlmutter, D.C. Cir., 18 March 2025) In England, the High Court found in November 2025 that Stable Diffusion's weights were not an infringing copy because they did not store Getty's images, a point Getty has permission to appeal. (Mayer Brown on Getty v Stability) The Munich court reached the opposite conclusion on lyrics because the lyrics came back out. The difference between the two cases is not the law. It is whether the work is reproducibly inside the model.

What this means for a performance model

RadioPro LLC builds AuralVerse, a text-prompted generative audio performance model. It renders lyrics and compositions that people have written and that the rightsholder brings to it. That design puts the model on the right side of each of the five conditions above, and outside most of the dispute the settlements resolve.

  • The work the model performs is licensed by definition. A composition reaches AuralVerse because its owner wants it performed. There is no training-time question about whether that song may be used, because the song is the input, not the corpus.
  • Identity is opt-in because there is no cloning. AuralVerse does not render in the voice of an identifiable living performer and does not accept prompts asking it to.
  • Outputs are controlled. The model is deployed inside RadioPro LLC's group of companies and affiliated networks. Renditions do not leave through a public download button.
  • Memorisation is tested against. The Munich reasoning sets a usable standard: a model that can be made to reproduce a recognisable recording it was not asked to perform contains that recording. We evaluate our models against that standard and treat failure as a defect, not a feature.

What this article does not do is describe how AuralVerse is built. The architecture, the weights and the training material are not published, and after this year's rulings we see no reason to change that. The courts in Munich did not ask OpenAI or Suno how their models work. They asked what came out. That is the question we think every model owner should be able to answer, and the only one that we answer in public.

The settlements of October and November 2025 did not end the argument about generative music. Sony is still litigating, the German judgments are under appeal, and the licensed platforms promised for 2026 have not yet launched. But the terms are written down now, and they will not loosen. Licensed input, opt-in identity, credited and paid participation, controlled output. A model that was designed that way from the start has nothing to renegotiate.