The Take — The Elvis estate is setting the price of a licence

The Elvis Presley filing is being read as a shakedown: a famous catalogue, a round number, a company that can afford it. I think it is something more useful and less comfortable. It is the first serious attempt to price AI training one song at a time — and the estates will get there before the record labels do, because the labels have already decided they would rather own a piece of this industry than litigate it into a per-work tariff.

That makes the $41 million suit a price signal rather than a damages claim, and it puts the number where the labs are weakest: not in what went into the model, but in what comes out of it.

Why an estate sues differently from a label

A label's business is licensing. Universal, Warner and Sony need generative AI to exist and to keep paying, which is why all three put money into Stability AI in August — two years after suing the company over models trained on their catalogues — in a $76 million round that came with deals to build music models from their own intellectual property. That is not hypocrisy; it is a business with a long catalogue and a long horizon deciding that an equity stake and a blanket rate beat a decade of litigation. We covered that flip when the round landed, and it is the clearest statement of how the labels intend to be paid.

An estate has none of that incentive. It owns a fixed catalogue that does not grow, a likeness, and a name that is worth more as a licensing line than as a platform play. Authentic Brands Group is not trying to shape the training-data market; it is trying to establish what one song costs. A label maximizing a catalogue measured in millions of works wants a blanket rate. An owner holding 274 songs wants a number with a comma in it.

That difference explains the shape of the complaint. The estate asked for $150,000 a song — the ceiling for willful infringement under the statutory damages provisions of the Copyright Act, against a $750 floor — as Rolling Stone reported, not a request for a licence. Read it as a demand: this is the price of one work, and there are thousands of us.

The complaint is built around the fair-use ruling, not against it

The reason the case leads with outputs rather than training is that the training argument has already been lost once, in the labs' favour. Judge William Alsup, in the Northern District of California, held that training Claude on copyrighted books was "quintessentially transformative" and protected as fair use; what Anthropic ultimately paid $1.5 billion for was a separate act — pulling more than seven million books from pirate libraries and keeping a central copy. We walked through that split when the settlement was approved, and it is the map every rightsholder lawyer now works from: attack the copy you can prove, not the learning you cannot.

So the estate attacks reproduction. The complaint's most quoted exhibit is a model response in which Claude, asked to rework specific songs "for a modern audience," complied and then described its own output as edging past "inspired by" into reproducing the copyrighted song.

That paragraph carries three counts at once. It is evidence of reproduction. It is evidence that a filter installed after this summer's lyric suits still leaks — Sony Music and Warner Chappell accused Anthropic of a "brazen campaign" of lyric piracy in a case we covered when it was filed. And it is evidence of knowledge, which is what turns a statutory floor into a multiplier. The estate is not arguing that Claude learned the songs. It is arguing that Claude will sing them, that Anthropic knew it, and that $150,000 is the statutory price of knowing.

Why the per-song number is the one that sticks

Here is the asymmetry that makes this more than a nuisance suit. A training claim is argued at the level of a corpus, where market harm is diffuse, transformative use is arguable and the numbers are enormous and vague. An output claim is argued one work at a time, and each one is a listening test: a jury does not need to understand tokenization to hear "Can't Help Falling in Love" come back out of a chatbot.

That matters commercially even if the estate never collects, because a blanket licence is not priced on what data is worth in the abstract; it is priced against the alternative, which is per-work exposure. The estate does not need a verdict to move the market. It needs a filing, discovery and a settlement number other estates can point at.

The second-order effect is what the open-weight camp should worry about. If per-work liability becomes a real cost of doing business, a legally clean corpus becomes a moat — exactly what Stability is selling when it markets models trained entirely on licensed music. Companies that can license works one at a time will do so and advertise it; everyone else inherits a liability it cannot price. That is copyright maximalism turning into a competitive advantage for the largest labs. Anthropic has committed $518 billion in infrastructure spending and can afford a catalogue. A graduate student redistributing weights cannot.

The counter-case, which is genuinely strong

Start with the law. Alsup's fair-use holding stands, and nothing in the filing changes it. Prompted reproduction is a different act with a different defence: the model was asked to rework specific songs, and a chatbot doing what its user instructed is not obviously Anthropic's business model. The company says the claims are wrong on the law and it will defend itself, it told The New York Times. Statutory maxima are reserved for conduct the defendant knew was wrongful, and Anthropic can point to a changing licensing posture, filters added under pressure and a good-faith record. Most copyright cases settle well below the demand.

Then the market argument. One estate is not a licensing regime. 274 songs is a rounding error against a training corpus, other estates will take blanket deals rather than courtrooms, and the labels — who own the breadth — have chosen equity plus licence. On that reading the filing is a bargaining chip in a negotiation that ends with a cheque and a non-disclosure, and the industry's real pricing mechanism stays where it was: a handful of blanket deals, signed quietly, between the largest rights holders and the largest labs.

And there is a version of this that is bad for artists: a per-work world rewards whoever can afford federal litigation. Elvis's catalogue can. The songwriter with forty songs cannot, and the estate route offers her no mechanism at all.

Why the take still holds

Because price discovery does not require a verdict, only a credible number. The estate has published one: $150,000 per work, in a public filing, attached to songs everyone knows. Every rights holder negotiating a training licence next year will walk in with it, and every lab will have to argue it down rather than argue from nothing.

And because the labs' real exposure is not the damages, it is discovery. A case like this is designed to produce an artifact — a work-by-work record of what a model can be made to reproduce, under oath, with exhibits. That is the document the licensing market has never had. Once it exists, in any case, in any district, the price of a licence stops being a negotiation and becomes an estimate.

What would change my mind

If the case settles for a nominal figure with a licensing deal attached — the Stability playbook, run in reverse — then the per-work theory was a bargaining chip and nothing more. If a court extends Alsup's reasoning from training to prompted reproduction, the willfulness count collapses back to the floor and the filing becomes a press release with a filing fee. And if Anthropic shows in discovery that its post-summer filters reliably refuse to reproduce the catalogue, the case reduces to a handful of bad outputs from a version of the model that no longer exists.

Any of those would make me wrong about the direction of travel. None would change the question the case has already put on the table, and which the industry has dodged since the first book was scraped: not whether AI may learn from a song, but what a song is worth once it can sing it back.

If a model can be made to perform a work you own, should the price be set per work or per corpus? Write to us via our contact page.