Elvis vs. Claude: Is AI Training Becoming the Music Industry’s Copyright War?

The King enters the AI courtroom

The dispute between music and generative AI has moved from an abstract argument about fair use into a harder question: did AI companies use valuable song catalogues as unlicensed fuel for systems that are now being sold through subscriptions, APIs and cloud platforms?

The latest trigger is Anthropic, the company behind Claude. On October 21, 2026, Bloomberg Law reported on a hearing in the Northern District of California in the long-running music-publishers case against Anthropic. The judge pressed both sides on whether the case could become one of the first major AI copyright trials over music training data. Days later, holders of rights connected to Elvis Presley songs sued Anthropic, alleging that Claude was trained on lyrics from 274 protected compositions and can generate infringing outputs. Bloomberg Law reported that the Elvis plaintiffs seek up to $48.6 million under copyright and DMCA theories; the Los Angeles Times reported roughly $50 million. Those are allegations, not findings. Anthropic told the Los Angeles Times it disagrees and will defend the case. (news.bloomberglaw.com)

This matters because the defendant is not a small experimental lab. Anthropic announced on May 28, 2026 that it had raised $65 billion at a $965 billion post-money valuation. Claude is sold not only through consumer subscriptions but through enterprise products, usage-priced API access and cloud distribution. Anthropic’s own pricing materials show per-token API billing, while Claude Enterprise is priced at $20 per seat per month plus usage billed at API rates. If copyrighted lyrics helped improve the model, the economic question is not symbolic: did culture become a cost-free input to a trillion-dollar infrastructure business? (anthropic.com)

The answer is not yet known. Lawsuits are accusations. Courts still have to decide what was copied, from where, for what purpose, whether Claude stores or reproduces protected expression, whether guardrails are meaningful, and whether fair use applies. But the cases reveal the central bargain still missing from the AI economy: if creative works are valuable enough to train models, why are creators so often told that permission and payment are unnecessary?

What is actually alleged?

There are now several overlapping music cases, but they are not identical.

The first major case, Concord Music Group and others v. Anthropic, began in October 2023 in Tennessee and was transferred to the Northern District of California, case number 5:24-cv-03811-EKL. The plaintiffs include major music publishers and allege that Anthropic copied protected song lyrics into Claude’s training corpus and that Claude could output protected lyrics. Anthropic’s position, as summarized in earlier court materials, is that including lyrics in training is fair use and does not require licences. (caselaw.findlaw.com)

That case has already produced useful judicial reality checks. In March 2025, the court dismissed secondary-liability and DMCA claims with leave to amend, while recognizing that the publishers had plausibly alleged Anthropic’s direct financial benefit theory for vicarious infringement. In October 2025, after amendment, the court allowed key amended secondary and DMCA claims to proceed. Separately, in denying a preliminary injunction, the court stressed that the requested order was too vague and sweeping: the publishers first pointed to 500 illustrative songs, but sought an injunction covering hundreds of thousands of works, including future acquisitions. The same order also said the publishers had not shown concrete evidence of lost licensing deals or renegotiated licences at that stage. That is crucial: the court did not bless Anthropic’s training, but it also did not accept broad market-harm claims without evidence. (loeb.com)

The Elvis case is newer. According to Bloomberg Law, the complaint was filed on October 23, 2026 in the Northern District of California and alleges infringement, circumvention of digital access controls, and removal of copyright-management information from about 274 works. The Los Angeles Times reports that the works include All Shook Up, Can’t Help Falling in Love, Don’t Be Cruel and Return To Sender. Available reporting identifies ABG EPE IP LLC and other rights holders, including publishing entities, trusts and individuals, rather than simply the Presley family. The complaint itself was not fully available to us through a directly reviewed court docket at publication time; therefore the most specific Elvis details in this article are attributed to Bloomberg Law, Los Angeles Times and other public reporting, not to our independent reading of every exhibit. (news.bloomberglaw.com)

Reservoir opened another front. AI Musicpreneur reported that Reservoir filed two Northern District of California complaints on October 21, 2026: one against Suno over 1,050 compositions and one against Anthropic over 853 lyrics, with maximum statutory-damages calculations of $186.4 million and $151.4 million respectively. Copyright Now reviewed the Suno complaint, identifying case No. 3:26-cv-12305 and describing a 28-page complaint plus a 23-page Exhibit A listing 1,050 works. For the Reservoir-Anthropic case, we treat the numbers as complaint-level allegations reported by specialist outlets; they have not been tested in court. (aimusicpreneur.com)

Claude is not Suno — and that distinction matters

Public debate often treats all AI music disputes as one story. That is sloppy. Claude is a text model. Suno is a music generator. A chatbot that outputs lyrics raises one kind of harm; a system that generates complete songs raises another. But both disputes meet at the same economic fault line: whether protected songs can be transformed into model capability without a licence.

The German GEMA/Suno decision shows why outputs matter. According to the European Commission’s IP Helpdesk, GEMA submitted 338 prompts relating to six works and obtained outputs it argued reproduced recognizable parts of the originals. The Munich court, case 42 O 763/25, focused on memorisation and largely upheld GEMA’s claims on July 31, 2026. Other legal summaries identify the works as including Atemlos durch die Nacht, Daddy Cool, Rasputin, Big in Japan and Forever Young; the full judgment and any appeal path should still be treated as jurisdiction-specific and not automatically transferable to U.S. fair use. (intellectual-property-helpdesk.ec.europa.eu)

Follow the money: who benefits, who pays?

The music side is not a romantic cottage industry. It is a global rights business. IFPI reported that global recorded-music revenues reached $31.7 billion in 2025, up 6.4%, with 837 million paid streaming subscription accounts. RIAA reported U.S. wholesale recorded-music revenue of $11.5 billion in 2025 and $6.4 billion from paid subscriptions. These figures cover recordings, not the full publishing market, but they show the scale of the sector AI is now colliding with. (ifpi.org)

Music publishing is the specific battlefield for lyrics and compositions. Reservoir’s SEC filing for the fiscal year ended March 31, 2026 shows total revenue of $175.7 million, with music publishing contributing about $116.8 million, or roughly 66%. Its publishing revenue came from digital, performance, synchronization, mechanical and other streams. That matters because AI training licences, if they become a real market, would likely sit beside these existing revenue categories rather than replace them cleanly. (sec.gov)

Concord is now part of a still larger consolidation story. In April 2026 BMG and Concord announced a combination; Bertelsmann said the combined company would generate pro forma 2026 revenue of $2.2 billion and EBITDA of $730 million. Concord said it had invested more than $3 billion since 2020 across publishing, recorded music, theatrical rights and distribution. This is not simply artists versus machines. It is also institutional capital, catalogue finance and global licensing strategy versus platform capital. (concord.com)

The Elvis case adds a different kind of ownership structure. Authentic Brands Group describes itself as a global brand and entertainment platform with more than 50 brands, 1,700 partners and over $38 billion in annual systemwide retail sales. UMPG announced in 2022 an exclusive global publishing deal with Authentic Brands Group to represent the historic Elvis Presley song catalogue. In other words, the public hears Elvis; the courtroom contains brand managers, publishers, trusts, estates and administrators. That does not make the claim illegitimate. It does mean that any settlement may enrich rights-owning entities before it reaches living writers or heirs. (corporate.authentic.com)

On the AI side, Anthropic’s business model is clear enough to analyze even without full public profitability. It sells access to Claude through subscriptions, enterprise seats, cloud marketplaces and token-metered API usage. The disputed legal point is whether training on lyrics is a legally permitted input. The economic point is simpler: better models attract customers, investors and cloud partners. If protected works improve a model, uncompensated training shifts value from copyright owners to AI firms and their investors. If licensing is required, value shifts back — but not necessarily to individual musicians.

The strongest AI argument

The AI side has a serious legal argument. U.S. copyright law has fair use. Courts have allowed copying for search, indexing, technical analysis and other transformative purposes. Anthropic also has a practical argument: if every work in every dataset must be individually licensed, only the largest firms and largest rights holders may be able to participate. A permission market could become a cartel of Big Tech and Big Rights.

The U.S. Copyright Office’s 2025 report does not give either side a clean victory. It says some AI-training uses may qualify as fair use and some may not. The analysis depends on what works were used, from what source, for what purpose, and with what controls on outputs. The report draws the line most sharply where expressive works are copied from pirate sources to generate unrestricted market-competing content when licensing is reasonably available; that, it says, is unlikely to be fair use. It also acknowledges the competition concern that mandatory licensing can entrench powerful incumbents. (copyright.gov)

That nuance is important. A model trained for scientific analysis is not the same as a model used to produce lyrics, music, stories or marketing copy. Licensed data is not the same as scraped or pirated data. A model with effective output controls is not the same as one that can reproduce protected text when prompted. Courts should not decide this by slogans.

The strongest music argument

The music side also has a stronger case than many technologists admit. Lyrics are not facts. They are compact, expressive and commercially licensed. If a model can output protected lyrics or produce close derivatives, the claim that training is merely internal technical learning becomes less persuasive.

The publishers’ best argument is not moral outrage. It is market substitution and bargaining harm. They argue that AI companies are building products that depend on expressive works while denying the existence of a licensable market. The weakness, as the court already noted in the preliminary-injunction phase of Concord v. Anthropic, is proof. General statements about lost control are not the same as evidence of a lost licensing deal, lower licence fees or measurable substitution. The publishers may yet prove those harms, but they have to do it with documents, data and examples.

There is another uncomfortable point: the music industry’s moral claim is strongest when it defends creators, but the money often flows through publishers, labels, estates, private companies and catalogue investors. A new AI-licensing market could become another opaque bulk-payment system. If songwriters receive little while catalogue owners negotiate large settlements, the result will be copyright enforcement without cultural justice.

Europe is choosing transparency first

The United States is still deciding fair use case by case. Europe has chosen a regulatory route. Under the EU AI Act, providers of general-purpose AI models must implement a copyright policy and publish a sufficiently detailed summary of training content. The Commission says these obligations entered into application on August 2, 2025, with enforcement powers applying from August 2, 2026; fines for GPAI providers can reach 3% of global annual turnover or €15 million, whichever is higher. (digital-strategy.ec.europa.eu)

This is not a full audit. A training-data summary is not a song-by-song disclosure. Trade-secret claims may hide important details. Small creators may not know how to reserve rights in machine-readable form. Still, Europe has accepted a principle the U.S. has not yet clearly imposed: AI developers owe rights holders and the public more transparency about the cultural material used to build general-purpose models.

What could go wrong next?

The first risk is a judicial green light for copy first, negotiate later. If courts broadly accept unlicensed training without meaningful transparency or output limits, AI companies will have little reason to negotiate except with the most powerful catalogues.

The second risk is the opposite: licensing becomes so expensive and centralized that only dominant AI firms and dominant rights holders survive. That would not liberate musicians. It would build a toll road between Silicon Valley and catalogue finance.

The third risk is evidentiary darkness. Without auditable training disclosures, creators must infer use from outputs, leaks, known datasets or discovery after suing. That is inefficient, expensive and structurally biased against ordinary songwriters.

The fourth risk is cultural flattening. Music is not merely content. It is memory, identity, labour and community. A platform economy flooded with cheap synthetic substitutes may give listeners abundance while reducing the bargaining power of human creators.

The evidence-based conclusion

The evidence supports three careful conclusions.

First, the music-AI conflict is no longer theoretical. There are now concrete lawsuits naming catalogues, works, outputs, damages theories and defendants. Second, the financial asymmetry is extraordinary: claims in the tens or hundreds of millions are being brought against AI businesses valued at hundreds of billions. Third, the law is unsettled. Fair use will turn on purpose, source, output controls, market harm, licensing availability and proof.

The unanswered question is whether courts and regulators will create a real market for permission, or legalize a retrospective model in which AI companies copy first, raise money, and negotiate only after litigation.

Elvis is a symbol, but this is not a nostalgia case. It is a test of whether human culture becomes paid raw material, unpaid raw material, or something even worse: a resource extracted so quietly that creators cannot even prove it was taken.