Can an AI learn from a song without competing with its writer? A judge hears both sides

A songwriter writes a line that becomes part of a copyrighted song. Years later, an artificial-intelligence system has absorbed large collections of text and can produce convincing lyrics in seconds. The legal dispute is not resolved merely by showing that the system learned language, nor merely by counting how many songs passed through training. In a California courtroom on 21 October, music publishers and Anthropic argued about an issue that may shape both AI training and the market for new creative work: when does using protected lyrics count as fair use?[1][2]

Reuters reported that Anthropic asked US District Judge Eumi Lee to hold that its use of song lyrics for model training is protected by the fair-use doctrine. Publishers including Universal Music Group, Concord and ABKCO argued that copying their protected material and the ability to produce competing lyrics threatened songwriter markets. The hearing involved competing legal interpretations and disputed evidence. It was not a verdict, and Judge Lee did not announce a ruling at the session.[1][2]

Training and output are two different questions

One question is whether it is lawful to reproduce protected lyrics as training material to build a general-purpose language system. Another is whether the system's subsequent output copies protected expression or competes in a way relevant to market harm. A court can evaluate these questions differently, because copying an existing lyric into a training corpus and generating a new lyric for a user do not necessarily involve the same conduct or the same asserted economic consequence. Treating the two as one slogan—AI steals music, or AI merely learns—discards the distinctions the arguments were about.[1][2]

Anthropic's position was that training is transformative: it uses the material to build a system performing a broader language capability rather than to redistribute the original collection of songs as such. The publishers' account emphasised the scale of copying and the ability to generate outputs that can compete with human lyric writing. These are positions put forward by adversaries, not findings this newsroom or a court has independently established.[1]

According to reporting on the hearing, the publishers pointed to nearly 120,000 requests for synthetic lyrics during a six-month period. The figure describes asserted user requests in that account, not 120,000 proven copyright infringements, completed songs sold commercially, or lost commissions to identified writers. The distinction is vital. Requests can help demonstrate demand for a feature, but translating them into a market-damage figure requires additional evidence about what was generated, what was used and what alternatives customers would otherwise have purchased.[1][2]

The hard part of market harm

Imagine two hypothetical customers. The first asks an AI to write a funny original song for a private birthday party. The second uses a system to reproduce a recognisable songwriter's protected lyrics for a commercial release. Both requests might appear in a broad product-use statistic, but their legal and economic significance would be radically different. The hearing forces attention onto how usage evidence connects to actual protected works and the market for licensing, substitution or downstream publication. The examples are illustrative, not incidents documented in this case.[1]

Copyright's fair-use analysis in the United States is contextual. Courts consider the purpose and character of use, the nature of the work, the amount used and effects on the potential market, among other case-specific considerations. Training-related litigation has produced competing judicial approaches, which is one reason the music dispute is being watched beyond the recording and publishing businesses. A result in one case would not automatically settle every model or every type of content.[1][2]

Bloomberg Law separately reported the judge's close questioning of both sides on market effects. That line of inquiry matters because the presence of a large training dataset does not settle how customers behave in the market. Equally, the absence of a perfectly itemised lost-sale list does not necessarily settle whether a licensing market can be affected. The parties are disputing what evidence is legally sufficient and how the framework should apply to a technology with both general and creative functions.[2]

What creators and AI builders can actually learn now

For publishers and songwriters, the case highlights two possible sources of bargaining power: control of the protected training material and control of the rights to reproduction or adaptation of the finished work. For developers, it raises the practical importance of data provenance, output safeguards, licence terms and records showing how content moves through the system. None of those measures alone guarantees a favourable court ruling; they are ways to make decisions more legible if those decisions are disputed.[1][2]

The lawsuit exists within a wider constellation of AI copyright cases, but the October hearing is a specific procedural event. Its significance comes from the competing arguments and the questions the judge is considering, not from a decision that has yet to be delivered. Reporting this accurately means keeping the unresolved state visible and refusing to treat either side's courtroom assertions as findings of infringement or exoneration.[1][3]

The larger argument is not whether machines can produce words that rhyme. They plainly can. It is who should have a claim when a commercially useful capability has been trained using other people's creative work, and what evidence shows that the new capability substitutes for something those people could otherwise sell. The answer will require law, economics and an unusually careful distinction between a dataset, an output and a market.[1][2]

Report an error or suggest a correction