Sony Music and Warner Chappell have filed a sweeping copyright lawsuit against Anthropic in the US District Court for the Northern District of California, alleging the company ran a "brazen campaign" of intellectual property theft by training its Claude models on tens of thousands of copyrighted songs without permission.

The damages being sought are eye-watering. The labels are claiming up to $150,000 per infringed work, plus up to $25,000 for every instance where copyright management information — the metadata that identifies who owns a song — was stripped from the training data. With "tens of thousands" of works named, the theoretical maximum runs into several billion dollars. That's not a number that gets thrown around for effect; in US copyright law, statutory damages at those levels are reserved for what courts treat as wilful infringement.

What makes this lawsuit different from the usual AI training scuffles is the specificity of the piracy allegation. Most AI copyright cases circle around whether training itself constitutes infringement — a genuinely unsettled legal question. This complaint reportedly goes further, accusing Anthropic of actively sourcing material from piracy sites rather than licensed datasets, which is a much harder position to defend. If the facts bear that out, the "fair use for training" argument that every AI lab has been sheltering under starts to look a lot shakier.

This isn't Anthropic's first brush with music copyright — Universal Music Group and others filed similar (if narrower) suits last year — but the Sony-Warner combination adds significant financial muscle and catalogue depth to the opposition. Between them, the two companies control catalogues that cover a substantial slice of recorded music history. That matters because the larger the catalogue, the higher the statutory damages ceiling climbs.

Why this matters beyond Anthropic. Every major AI lab has trained on internet-scale data scraped without explicit licensing agreements. OpenAI, Google DeepMind, Meta — all face versions of this question. What's different here is that if a court finds Anthropic specifically sought out pirated sources, it creates a factual distinction that could expose the company to damages that dwarf anything seen in AI litigation so far. A settlement, whenever it comes, is likely to set a de facto licensing rate that the rest of the industry will have to match.

The Australian angle is real. Australian copyright law doesn't have the same statutory damages regime as the US, but the government's ongoing review of the Copyright Act specifically flagged AI training as an unresolved area. If US courts establish that unlicensed training is infringement (rather than fair use), Australian AI developers and researchers operating under the assumption that training is implicitly permitted would need to recalibrate fast. ARIA and APRA AMCOS have both been watching this space closely.

Watch for Anthropic's initial response — whether they contest the piracy-sourcing allegation head-on or try to shift the argument back to fair use will tell you a great deal about how strong their internal evidence actually is.