The Copyright War That Courts Are Losing

In March 2025, Judge Arun Subramanian of the Southern District of New York ruled that Suno Inc.’s training of its music generation model on copyrighted recordings without license constituted copyright infringement. Three months later, a different judge in the Northern District of California reached a different conclusion on substantially similar facts in the UMG v. Udio case, finding that the degree of actual reproduction was insufficient to constitute infringement. By early 2027, there are nine active appellate proceedings in US federal courts dealing with some version of this question, and the circuits are not aligned.

This is what a legal system looks like when it is genuinely unprepared for a technological shift. Not obstructionist, not captured by industry lobbying — simply unprepared, working with doctrinal tools designed for a world where copying meant copying, where training a model on a song was not a category the Copyright Act of 1976 needed to address.

What the Cases Actually Argue

The central legal question in training-data copyright litigation is whether using copyrighted works to train a machine learning model constitutes reproduction in the legally meaningful sense. The answer depends on how you characterize what training does.

One view: training is transformative use. The model that results from training on a million songs does not contain those songs; it contains statistical representations of patterns in those songs. Using a work to train a model is analogous to a musician listening to thousands of records and developing their own style — nobody owns style, and nobody can copyright the influence their music had on the next generation of artists.

The competing view: training is reproduction. To train a model, you must copy the works into computational infrastructure. That copying is reproduction. The eventual outputs of the model are less relevant to this question than the copying itself, which occurred without license. The transformativeness of the downstream use does not retroactively license the upstream reproduction.

Both views are defensible under existing doctrine, which is the problem. The fair use doctrine in US copyright law — which allows unlicensed use of copyrighted material under certain conditions — was developed through cases involving photocopiers, television recorders, and internet search engines. The Supreme Court’s 2023 Andy Warhol Foundation decision, which found against Warhol’s estate in a licensing dispute, emphasized the commerciality of the downstream use as a critical fair use factor. That emphasis has complicated AI training arguments, since the training unambiguously serves commercial purposes.

The European Approach

The EU took a different route through the Text and Data Mining exception in the Digital Single Market Directive of 2019. Under this exception, using protected works for text and data mining is permitted, with an opt-out provision — rights holders can declare their works off-limits for commercial TDM purposes.

The music industry in Europe moved aggressively on opt-outs. By early 2026, the three major labels (Universal, Sony, Warner) had issued blanket opt-out declarations for their catalogs, as had several collecting societies representing independent labels. Whether these opt-outs are technically effective — particularly for AI systems trained before the opt-outs were issued — is itself a matter of ongoing litigation in Germany and France.

The EU framework, despite its imperfections, has at least produced a mechanism for negotiation: licensing deals rather than purely adversarial litigation. Universal Music Group signed a licensing agreement with one AI music platform in late 2024, terms undisclosed, that was widely seen as a template for industry accommodation. Whether the amounts involved are commercially meaningful to either side is unknown, but the existence of a licensing pathway is different from the US situation, where no established framework for licensing training data has emerged from litigation.

What the Rulings Don’t Address

The legal battles are focused almost entirely on training data — whether the unauthorized use of copyrighted works to train models is infringement. They are saying relatively little about the output side: whether AI-generated music that sounds similar to a specific artist’s work infringes that artist’s copyright or violates their right of publicity.

This matters because the training question, while economically significant, is not where the consumer-facing harm is concentrated. The more visceral harm to working musicians is not that their albums were used to train a model they didn’t license — it is that the model now generates music that sounds like them, that competes with them for placements in advertisements and film scores and streaming recommendations, and that is produced at zero marginal cost.

Sound recordings are protected by copyright. Musical compositions are protected by copyright. An artist’s particular style — the tonal quality of a specific voice, the idiosyncratic rhythmic approach of a specific producer — is not protected by copyright. It never was. The principle that style is not copyrightable predates AI by a century. What AI has done is create systems that can execute style imitation at industrial scale, which converts a theoretical legal freedom (imitate but don’t copy) into a practical competitive threat (produce unlimited content in the style of any artist, at minimal cost, with no royalty obligations).

The Soundalike Economy

There is already a substantial market for AI-generated music that operates on the style-but-not-copy logic. Several production music libraries — the companies that supply background music for YouTube videos, advertisements, podcasts, and corporate presentations — have transitioned partially or fully to AI-generated content. The pricing is roughly 80% below what licensed production music previously cost. The quality, for ambient and background applications, is sufficient for most purposes.

This has not collapsed the commercial music market. It has fragmented it. The market for music in professional production contexts is bifurcating: the high end (sync licensing for major film and television, major advertising campaigns, live performance) remains expensive and human-created, and the value attached to human creation is explicitly marketed. The low end (social media content, corporate backgrounds, YouTube filler) is largely AI-generated or becoming so.

What has collapsed — partially but consequentially — is the middle market. The production music library that previously paid individual musicians reasonable rates to produce thousands of tracks for stock use is no longer a viable economic model. This is not hypothetical. AudioJungle’s transaction volume for human-created stock music dropped approximately 60% between 2023 and 2026. Pond5, Artlist, and similar platforms have shifted toward licensing AI-generated music at dramatically lower prices. The musicians who depended on this market as a supplement to other income have lost it.

The courts are adjudicating the wrong question, on the wrong timeline, with the wrong tools. The training data question will eventually be resolved — probably through a combination of legislation, licensing frameworks, and appellate decisions that create clearer precedent. By the time it is resolved, the economic structures it is trying to govern will have evolved past the point where the resolution applies cleanly. This is not unique to AI copyright. It is the normal condition of technology law. It is just moving faster than usual.