Ask a large language model to write something in the manner of a well-known songwriter, and you get one of two results. Either bland pastiche, or a verbatim block of lyrics that somebody still collects mechanicals on. The second outcome is a screenshot, and screenshots are what turned a diffuse industry anxiety about training data into pleadings with exhibit numbers. That is why the Anthropic copyright litigation is worth reading as a set rather than as a headline: one defendant, two very different plaintiff groups, and a third case next door against the audio generators. Three fact patterns, three theories of where the harm attaches, and by now enough procedural history to compare them on something firmer than sentiment.
What the Anthropic copyright litigation is actually about
Two separate questions that most coverage runs together. The first is acquisition: how the training corpus was assembled, and whether the copies inside it were obtained lawfully. The second is output: whether the deployed model hands protected expression back to a user. A defendant can win the output question and still lose the acquisition question expensively. That is approximately what happened in the authors' case, Bartz v. Anthropic, where Judge William Alsup held in June 2025 that training on lawfully purchased books was transformative, while retaining downloaded pirate-library copies in a permanent internal collection was not. The settlement reported at roughly $1.5 billion was priced off the second holding, not the first.
Criterion one: where the harm attaches
The music publishers' suit — Concord Music Group and co-plaintiffs, filed in Tennessee in late 2023 and later moved to the Northern District of California — is built on output. Its exhibits are model responses containing lyrics to compositions the plaintiffs administer, alongside the argument that the copies made to train were themselves infringing. The pleading has been amended more than once as the case progressed, with later versions drawing on material produced in discovery about how the corpus was handled internally. Those characterisations belong to the plaintiffs and remain contested.
The authors' case barely needed the output at all. Its centre of gravity was procurement: what was downloaded, from where, and what was done with it afterwards.
The label suits against Suno and Udio, filed in mid-2024 with RIAA coordination, sit between the two, but their sharpest claim is output — generated audio said to be substantially similar to identifiable masters.
Criterion two: how cheaply the evidence reproduces
This is the criterion that separates them, and it gets underweighted.
A lyric claim reproduces for free. Anyone with a browser runs the prompt, and either the protected string comes back or it does not. That is enormously persuasive right up until the defendant ships a filter, at which point the demonstration stops working and the claim turns historical. Reporting through 2025 described a stipulation on maintaining output guardrails, followed by the court declining to grant the publishers a preliminary injunction on the broader training claims. By then the live demo had already been closed.
An audio claim never reproduced cheaply in the first place. Substantial similarity in recorded music needs musicologists, and "sounds like a 1968 Motown rhythm section" is not protectable. Genre is not owned. That expert layer costs money before a single fact is established.
An acquisition claim does not decay. Download logs, procurement decisions, and internal correspondence are fixed in the past and largely produced by the defendant. No product change unmakes them.
Criterion three: whose paper is clean
Books have tidy plumbing: one title, one registration, usually one owner. Compositions do not. A single song can carry co-writers, split shares, sub-publishers, territorial carve-outs, and an administration agreement that determines who is even entitled to sue. Masters add another layer of consents. The cleaner the chain of title, the faster a case reaches a number — which has nothing to do with the moral weight of the claim and everything to do with how a class gets certified.
The comparison, in one table
| Publishers v. Anthropic | Authors v. Anthropic | Labels v. Suno / Udio | |
|---|---|---|---|
| Harm attaches to | model output | corpus acquisition | output, plus acquisition |
| Reproducible by a non-expert | yes, until filters ship | not required | no |
| Ownership plumbing | splits, co-writers, admin deals | one registration per title | masters plus artist agreements |
| Fair use pressure point | substitution for licensed lyric display | source of the copies | substitution for licensed catalogue |
| Status as of writing | live, narrowed by rulings | settled | moved toward licensing deals |
The thing the three cases have in common
None of the resolutions so far came from a court deciding what a model may generate.
The books matter settled on how copies were obtained. The audio matters, as of writing, moved toward licensing arrangements rather than judgments, which leaves the doctrinal question — whether a generated recording that evokes a master infringes it — untested. The publishers' claims still standing are the ones that reach behind the interface.
If you are modelling exposure, that ordering is the finding. Output claims are visible, quotable, and perishable. Acquisition claims are invisible, unglamorous, and permanent. The document that decides these cases is usually not a transcript of the model talking. It is a receipt.
Everything downstream inherits that. When we compare generators' licence terms on this site, the clause that moves is rarely the one about what you may do with your render. It is the one describing what the vendor asserts about its own inputs, and how quietly that assertion can be revised.
Since the books ruling landed, there is a plain text file sitting beside the stems in every session folder on my drive: model and version, the prompt verbatim, the date I pulled the licence terms, and where any reference audio came from. It has never improved a mix. It is the only part of the folder I could hand to a lawyer.
Not sure which tool to use?
Compare the top AI music and sound tools side by side — honest reviews, real pricing, no sponsorships.
