Sony's Anthropic Lawsuit Changes the AI Copyright Game Forever

The Real Target Isn’t Just Anthropic

Thirty-five music publishers, led by Sony Music Publishing, filed a federal lawsuit in California on August 28 alleging that Anthropic built Claude by training its models on tens of thousands of copyrighted lyrics and sheet music without permission or payment. The complaint names CEO Dario Amodei and the company’s co-founders individually.

The case lands at a moment when the broader AI-copyright conversation looks settled — or so it seemed after Anthropic agreed to pay approximately $24 billion in late July to settle a lawsuit brought by book authors. The terms of that deal required Anthropic to confirm that its training data included pirated copies of millions of books. The court approved the settlement. Fair use for text-based AI learning was effectively confirmed, with the narrow exception of illegal sourcing.

The music lawsuit traces the same legal path. But something about the timing and the plaintiff mix suggests this was never just about compensation. It was always about establishing a different legal posture — one where the question isn’t whether Anthropic obtained its data through piracy, but whether the act of training itself constitutes infringement.

Why This Case Is Different

The legal distinction matters more than it first appears. In the authors’ settlement, the core finding was that learning from books is permissible under fair use; the violation was that the books came from pirate sites. The court never ruled that AI training on legitimately obtained text is unlawful.

The 35-label complaint goes further. Music publishers are arguing that the training process itself — ingesting copyrighted musical works to produce derivative AI outputs — is the violation. This is a harder line to draw, but it aligns with language from the ongoing Meta case before the same federal court, where the judge suggested that large-scale unauthorized training on creative works may constitute infringement regardless of how the data was acquired.

That Meta case is currently on appeal. Whatever ruling emerges from that docket could define the legal standard for every generative-AI company in existence.

The Stakes Are Structural

If Anthropic loses, the implications extend far beyond one company. Every major AI platform — OpenAI, Google DeepMind, Microsoft, Perplexity, others — trains its models on creative content scraped from the internet. Music lyrics are among the most commercially valuable forms of copyrighted material on the web. A ruling against Anthropic on the music-training theory would effectively require every AI developer to negotiate licenses with every rights holder whose work appears in their training data.

That is not a practical roadmap. The music industry alone manages millions of works across thousands of publishers and collecting societies. Licensing every track, lyric, and composition at scale would consume enormous resources and delay product timelines indefinitely. Many smaller AI startups would simply be unable to comply.

The alternative path — paid licensing databases — is emerging but remains incomplete. Music licensing deals exist in fragments: certain publishers have struck agreements, certain datasets are available for purchase. But there is no unified licensing infrastructure for music training data comparable to what exists for text or code. Building one would be a multi-year project with no guaranteed commercial model.

The Japanese Angle

Japan offers a useful counterpoint. In 2024, Japan revised its copyright law to permit AI training on copyrighted works as a general rule, with limited exceptions. The law was designed to make Japan a competitive location for AI development while preserving creators’ rights through opt-out mechanisms and collective management frameworks.

The revision did not settle the fundamental question: whether training on copyrighted works should be free, subsidized, or fully licensed. Japan chose permissiveness. The United States is choosing adjudication. The two approaches will produce different industry structures, and the companies that operate across both jurisdictions will face conflicting obligations.

For Sony, the choice is clear. Japan is one of the world’s largest music markets. Sony Music Publishing represents one of the deepest catalogs in the industry. Allowing AI companies to use Japanese and global musical works without compensation undercuts Sony’s core business — licensing and publishing revenue that funds artist development, catalog maintenance, and creative investment.

Who Wins and Who Loses

If Sony and the 35 labels prevail, the winners are established music publishers and legacy artists with deep catalogs. The losers are AI developers who relied on scraped creative content as a free input, and the startups that would have used that content to compete. The structural advantage shifts toward incumbents who own creative works and can afford licensing deals.

If Anthropic prevails, the winners are AI developers and the broader tech ecosystem that depends on open creative data. The losers are music publishers who see their catalogs treated as public training resources without compensation. Startups benefit too, since licensing friction falls on incumbents rather than newcomers.

The current trajectory favors Anthropic. The authors’ settlement already established that fair use covers AI training on legitimately sourced text. The music case adds a new theory — that training itself is infringing — but the same federal court that approved the settlement has not signaled openness to that argument. The Meta case remains the live variable.

What Comes Next

The lawsuit will proceed through discovery, likely extending into 2027. Anthropic has every incentive to settle early rather than risk an adverse ruling that would bind the entire industry. The authors’ settlement pattern suggests a payments-based resolution, not a precedent-setting judgment. But the music industry has more bargaining power than the authors did — not just because of Sony’s catalog depth, but because 34 other publishers are joined as plaintiffs, creating a coalition that could coordinate future litigation or licensing demands.

The outcome will determine whether AI companies must treat creative works as inputs to license or as freely available data. That decision reshapes the economics of the generative-AI industry. Everything after this lawsuit depends on it.