Music – Part 1: Litigation & licensing

Introduction

AI is changing how music is made and how the business works. Artists are using it in production, developers are negotiating catalog licenses, and streaming services are deciding how to treat AI-generated tracks. Disputes over training, imitation, and payment are testing the limits of existing law.

For artists and music companies, the questions are practical: What uses require permission, who can give it, and what rights protect the resulting music? The answers affect recording agreements, catalog values, licensing terms, and royalty income. AI also offers ways to address longstanding problems in rights administration and collections. This article examines those legal and commercial developments and the decisions facing the industry.

A central question in AI copyright litigation is whether copying protected works to train a model without permission constitutes infringement or qualifies as fair use. For music, that inquiry can involve separate rights in sound recordings and the underlying compositions, as well as distinct acts of copying during acquisition, storage, and training.

The fair use debate

Generative AI models are trained on large datasets that may include copyrighted works obtained without the rightsholders’ permission. Rightsholders argue that this copying infringes their exclusive rights and allows developers to build competing products using their catalogs. Developers contend that training serves a different purpose from the original works and is protected by fair use.

Fair use is an affirmative defense to copyright infringement. Under 17 U.S.C. § 107, courts assess four factors: (1) the purpose and character of the use, including whether it is commercial; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used; and (4) the effect on the potential market for or value of the work. Whether a use is transformative is an important part of the first factor, but does not by itself resolve the analysis.

In the training cases, much of the dispute concerns the relationship between the first and fourth factors: whether using works to develop a model serves a sufficiently different purpose, and how to account for the model’s ability to generate content that competes with those works.

The courts

In June 2025, the U.S. District Court for the Northern District of California issued the first two decisions on the merits of fair-use defenses to copyright infringement claims arising from generative AI training: Bartz v. Anthropic and Kadrey v. Meta. Both judges held that the training uses before them qualified as fair use and therefore did not infringe the plaintiffs’ copyrights, but their reasoning differed, particularly on market harm.

In Bartz, the court found that using the plaintiffs’ books to train Anthropic’s language models was highly transformative. It separately rejected Anthropic’s attempt to justify its acquisition and retention of pirated books for a permanent library on that basis. The training ruling did not excuse the separate library copying.

The distinction had substantial financial consequences. Anthropic subsequently agreed to a $1.5 billion settlement following the ruling that left the piracy claims for trial. The settlement did not reverse the court’s fair-use ruling on training.

On market harm, Bartz rejected the argument that an increase in competing, noninfringing works was itself a harm copyright law protects against. Its reasoning went beyond finding the plaintiffs’ evidence insufficient: It rejected that form of competitive displacement as a basis for weighing against fair use.

Kadrey took a different approach. Although the court also found training transformative, it recognized that a model’s ability to generate large quantities of competing works could dilute the market for the works used in training. Meta nevertheless prevailed because the plaintiffs had not adequately developed the arguments and evidence needed to support that theory. The court emphasized the limits of its ruling and did not give blanket approval to unlicensed training. We discussed these decisions in A New Look at Fair Use: Anthropic, Meta, and Copyright in AI Training.

In music, pro rata streaming royalties give the indirect market-harm theory a concrete financial dimension. Under that model, recordings compete for a share of a finite royalty pool in each accounting period, so streams of eligible AI-generated tracks can reduce the share paid to existing recordings even when the AI outputs do not infringe any particular work. Whether that dilution constitutes harm recognized by copyright law, and what evidence would connect it to the challenged training, remain unresolved. The differing approaches in Bartz and Kadrey leave those questions very much alive.

The government

The U.S. Department of Justice has also weighed in. In a September 2026 filing in the New York litigation involving The New York Times, OpenAI, and Microsoft, DOJ supported the fair-use defense for the training at issue. It emphasized the public benefits of AI development and argued that restrictions on training could impair U.S. economic competitiveness and national security.

The U.S. Copyright Office has taken a more qualified position. In Part 3 of its Copyright and Artificial Intelligence report, the Copyright Office concluded that some training uses will qualify as fair use and others will not. It expressed particular concern about copying expressive works from pirate sources to generate competing content where licensing is reasonably available. Its analysis considers how works are acquired, the purposes of training, the capabilities of the resulting model, and the effects on relevant markets.

Neither the DOJ filing nor the Copyright Office’s report binds the courts. They nevertheless offer different assessments of how copyright law should apply to AI training and how much weight to give the technology’s asserted benefits relative to its effects on creators and licensing markets

Licensing: A parallel track

Music companies and AI developers are negotiating licenses while continuing to litigate. Warner Music Group’s November 2025 agreement with Suno, for example, settled the litigation between them and announced a partnership involving licensed models and artist and songwriter participation.

Consent is a central industry demand, but whose consent counts remains disputed. Labels and publishers insist that AI developers obtain permission to use their catalogs, yet artists and songwriters do not always have an unconditional right to keep their music out of those deals. In a June 2026 statement, 29 creator and manager organizations alleged that music companies were licensing works without consulting their creators, sometimes seeking consent for voice uses but not training. Whether existing contracts authorize those licenses is a separate question from whether creators should have a say.

The deals also bear on the fair-use disputes. Rightsholders point to negotiated licenses as evidence of a market for AI training rights and argue that unauthorized training deprives them of licensing revenue. Developers respond that payments may reflect settlement risk or a broader commercial relationship, rather than the value of training rights alone.

The existence of a paid license does not, by itself, establish that the challenged use requires one. Nevertheless, the Copyright Office’s report treats available or reasonably developing licensing markets as relevant to the market-harm analysis. The scope of each deal matters: A payment covering past claims, future training, voice uses, and output exploitation may reveal little about the value of training permission in isolation.

Restrictions on outputs are another significant negotiating point. Under a “walled garden” approach, generated tracks remain within the AI service rather than being exported for distribution on streaming platforms. Other arrangements permit downloads subject to limits. The Warner–Suno announcement, for example, contemplated downloads for paid subscribers, with monthly caps, while restricting downloads from the free tier.

For rightsholders, those restrictions address whether licensed AI tools will generate music that competes with their catalogs on streaming services and for other licensing opportunities. The commercial bargain therefore extends beyond permission to train: It also concerns where generated music can be used, how artists and songwriters are paid, and what control they retain.

Where things stand

• The early training decisions favor developers, but do not establish a general exemption. Both rulings turned on the uses and records before the courts, and their approaches to market harm differed. They are persuasive authority for other courts, rather than binding precedent.
• Indirect market harm remains a significant issue for music. Pro rata streaming royalties provide a concrete setting for assessing competition from AI-generated tracks. Whether resulting royalty dilution constitutes legally cognizable harm, and what evidence would establish it, remain unresolved.
• A favorable training ruling does not resolve output claims. Whether a particular output infringes protected expression, and who may be responsible, requires separate analysis.
• Licensing leaves important questions of consent and control to negotiation. An AI license from a label or publisher does not necessarily mean that each affected artist or songwriter has separately consented. Training rights, output restrictions, compensation, and creator approvals all need to be addressed.