Claude Lawsuits

The training-data copyright theory

The dominant first-wave theory against generative-AI labs has been straightforward: training a large language model requires ingesting tens of billions of words, the cleanest sources of high-quality text are copyrighted books and articles, and copying those works into a training corpus — even temporarily — is reproduction within the meaning of the Copyright Act. Plaintiffs argue the training itself is therefore an infringing use; the labs respond that training is transformative fair use under Authors Guild v. Google (the Google Books decision) and Sony v. Universal (the Betamax decision).

The theory shows up in Bartz v. Anthropic (books), in Concord Music Group v. Anthropic (song lyrics), and in the OpenAI docket on the GPT side (NYT v. OpenAI, Authors Guild v. OpenAI). What's distinctive about the Anthropic docket is that it produced the first federal ruling on the merits — the Alsup summary-judgment opinion in Bartz — before any of the OpenAI cases got past the pleading stage.

The Alsup ruling and the piracy distinction

On June 23, 2025, Judge William Alsup of the Northern District of California granted partial summary judgment in Bartz v. Anthropic. The opinion split the copyright question along a line that had been theoretical until the ruling landed and is now load-bearing: training versus acquisition.

On training, the court held that running a model over lawfully-acquired text to learn statistical patterns is transformative fair use. The model does not output the underlying works (Claude does not, on demand, recite a Bartz novel verbatim), the use is "spectacularly transformative" relative to what the books are for, and any market-harm theory has to be grounded in something more than the speculative claim that a more-capable Claude makes book sales harder.

On acquisition, the court held that Anthropic's downloading of pirated book copies from sites including LibGen to build a permanent in-house corpus is not fair use, regardless of what the corpus is later used for. The acquisition is itself the infringing act — the same way that buying a stolen book is illegal regardless of whether you later read it for a permitted purpose. That holding teed the case up for a damages trial covering more than seven million books Anthropic had pirated during corpus construction.

The line the Alsup ruling drew — train freely on what you have legitimate access to; do not source through piracy — is the most-cited single passage in LLM-copyright law as of early 2026. It is the lodestar every subsequent training-data complaint and answer reads against.

The Bartz settlement

Rather than try the piracy-damages question, Anthropic settled in August 2025 for $1.5 billion — the largest copyright settlement in U.S. history by a wide margin. The structure is four installment payments running through September 2027, distributed at roughly $3,000 per eligible work to authors of the approximately 500,000 books in-scope at settlement.

The settlement-administration vehicle at anthropiccopyrightsettlement.com is the authoritative source for the claim mechanics: who qualifies, how to file, how the per-work amount is calculated, and how the four-installment schedule plays out. The claim deadline was March 30, 2026.

Two things the settlement notably does not do. It does not undo or vacate the Alsup ruling on training fair use — that part of the opinion stands and is now precedent. And it does not require Anthropic to delete the trained-model weights; the settlement is about compensation for the pirated-acquisition stage, not the training output. Both points are deliberate.

Music-publisher coordination — the Concord case

Concord Music Group, et al. v. Anthropic was filed in October 2023, predating the broader wave of LLM-copyright litigation. The plaintiffs — Concord Music Group, Universal Music Publishing Group, and ABKCO Music — are major music publishers that hold rights in song lyrics. The complaint runs on two tracks: a training-input claim (Anthropic ingested copyrighted lyrics without licensing them) and an output-reproduction claim (Claude, when prompted for the lyrics to specific copyrighted songs, returns the lyrics verbatim or near-verbatim).

The output-reproduction track is the more novel of the two. Bartz-style training claims have been heavily litigated; output-side claims have been harder to bring against general-purpose LLMs because the models do not, in the ordinary case, emit copyrighted long-form text. Lyrics are an exception: short, well-known, and within the model's memorization window. A partial preliminary injunction was entered early in the case directed at the model's lyric-emission behavior — Anthropic-side guardrails were tightened in response.

Venue moved from the Middle District of Tennessee toward the Northern District of California, where the broader AI-copyright bar is concentrated. The case remains active. Whether the Alsup training-fair-use line carries over to music compositions, and whether the output-reproduction theory survives summary judgment, are the two questions to watch.

Platform data licensing — the Reddit theory

Reddit v. Anthropic, filed in June 2025, runs on a different track from Bartz and Concord. There is no copyright claim. Reddit's theory is contract: Anthropic accepted Reddit's terms of service when it accessed Reddit content programmatically, those terms forbid bulk training-data scraping without a paid license, and Anthropic continued scraping after Reddit's licensing program demanded that scrapers either pay (as OpenAI and Google did) or stop.

The complaint pleads breach of contract, unjust enrichment, and unfair competition under California's Unfair Competition Law (Cal. Bus. & Prof. Code § 17200). The procedural fight since filing has centered on whether the case stays in California state court (Reddit's preferred venue, where the UCL has more bite) or moves to federal court (Anthropic removed; Reddit moved to remand). The substantive merits questions are paused on that procedural posture.

Why this case matters separately from the copyright cases: contract liability is not reachable by fair-use defenses. If Reddit prevails on the breach-of-contract theory, every AI lab that crawled a major social platform has parallel exposure under the same theory, regardless of how the underlying training-fair-use question resolves. Reddit v. Anthropic is the leading test of that proposition. Reddit has separately signed paid-licensing deals with several other AI vendors; the litigation is leverage as much as a damages claim.

What this docket means for the broader AI bar

Anthropic's three lawsuits between them touch every major flavor of AI training-and-operation theory: copyright on the training input (Bartz, Concord), copyright on the output (Concord's lyric-reproduction track), and contract on the platform-licensing question (Reddit). The Alsup ruling resolved the training-input question on the books side in the labs' favor and resolved the piracy-acquisition question against them. The other two cases are testing whether those distinctions hold up across the music and platform contexts.

The broader effect, as of early 2026: the LLM-copyright bar is concentrating in the Northern District of California, fair-use defenses on properly-acquired training data are stronger after Alsup than before, piracy-sourced corpora are uniquely exposed, output-reproduction theories are alive but not yet tested at trial, and the contract-and-TOS theory is the open frontier.