Elvis's rights holders put a price on AI training: labs should license music

What is more dangerous to an AI lab: a vague argument about whether training is fair use, or a list of 274 song titles with a price tag attached? The Elvis lawsuit against Anthropic suggests the second, and we think that should change how labs behave.

Here is the short version. Authentic Brands Group, through a unit called ABG EPE IP LLC, and co-plaintiffs including songwriter trusts and the publishers Abilene Music and Raleigh Music Publishing, sued Anthropic on October 23 in federal court in Northern California. As Infobae reports, they allege Claude was trained on 274 Presley songs without permission. They seek $41.1 million, which is the $150,000 maximum multiplied by 274 works. They add about $7.5 million for circumventing access controls and stripping rights information, plus destruction of the unauthorized copies and a jury trial. Anthropic had not responded publicly at the time of reporting.

The countable work is the new unit of risk

The abstract fair-use fight is hard to price. A catalog is not. Once a plaintiff can name 274 specific works, the statutory-damages math is just arithmetic. That is why the $41 million figure matters more than its size. For a company of Anthropic's scale it is probably not existential, but it is a repeatable template. This is the latest in a long line of music suits against Anthropic, including those from Universal Music Publishing, Concord and BMG, and a Reservoir complaint over 853 compositions.

Multiply that template by every catalog in the world and the risk stops being a legal curiosity. It becomes a recurring cost, whether or not it ever reaches a verdict. In our view, that is the point at which finance teams, not lawyers, start asking for licenses.

Winning on training may not save you

Labs often lean on the argument that training is fair use. It may even win. But as Mubert's analysis notes, citing Forbes, Judge William Alsup ruled in 2025 in Anthropic's books case that training on copyrighted text could count as fair use, while storing millions of pirated books in a permanent library did not. The Elvis complaint also puts weight on how the data was obtained, including the alleged circumvention of protections on licensed platforms.

That is a weaker position than "we transformed it." You can win the philosophical question and still lose on sourcing. We would not bet a roadmap on that distinction.

The most awkward exhibit is the chatbot itself

The complaint quotes a response in which Claude, asked to recreate Presley songs for a modern audience, allegedly said the results stayed very close to the originals, with the same structure and reused hooks. It also allegedly said they were closer to reproducing the protected song than to being "inspired by" it. We should be careful here: this is the plaintiffs' account of an output, and a court will test it. But as a public-relations problem, it is brutal. The plaintiffs use it to argue that the guardrails Anthropic says it adopted after earlier suits are not enough.

And then there is the name. Juries know who Elvis is, and most have never heard of ABKCO, as Mubert puts it. The legal theory is not new, but the face is.

What labs should do

Our recommendation is plain: license. Not because the law has been settled against them, since it hasn't, but because litigation across dozens of cases is a slow way to buy certainty. A path is already visible: Suno reached a settlement with Warner Music Group last year while still fighting Universal and Sony Music. Vice also notes that disputes like these often end in settlements and quiet partnerships.

There is a fair counterargument. Licensing every catalog could entrench the biggest players and lock out smaller labs and open projects. That is a real cost, and a reason to push for collective or standardized licensing rather than one-off deals. But "it would be hard" is not a defense, and rights holders are now showing they will count song by song.

The verdict: the Elvis suit adds no new doctrine, but it makes the price visible. A risk you can put in a spreadsheet is a risk you can buy out. Labs that start negotiating now will do it on better terms than those that wait for a jury in a room full of people humming "Can't Help Falling in Love."