Sony Music Publishing and Warner Chappell have filed a 48-page lawsuit against Anthropic, naming its founders personally, over the use of song lyrics in training its models. The publishers are seeking up to $150,000 per work. The amount is no accident: it is the US legal cap for wilful infringement, and a precedent at $1.5 billion already exists.
This year we have followed the book publishers' lawsuits, the question of generated music, and the awaited verdict in the GEMA case in Europe. Here is the next front, and it is better prepared than the previous ones.
Why this lawsuit is different
Three elements set this action apart from those that came before it.
The precedent exists. Anthropic settled a dispute with book authors last year for a reported amount of around $1.5 billion. Music publishers therefore know the company has already agreed to pay, and at what level. A lawsuit filed after a settlement is a lawsuit that knows the price.
The founders are targeted personally. This is unusual and strategic. Naming individuals as well as the company increases the pressure and makes the case harder to treat as a mere operational cost.
The amount aims for the cap. $150,000 per work corresponds to the maximum provided under US law for infringement deemed wilful. Claiming that cap is a way of asserting the company knew.
What makes lyrics distinctive
Song lyrics hold a special place in this debate, and it is worth understanding why.
A book contains thousands of sentences, none of which a model reproduces in full. A song contains a few dozen lines, often repeated, extremely prevalent on the web, and which a model can output word for word. The question of memorisation, which we raised regarding test contamination, arises here directly: did the model learn to write, or did it learn the song?
That is what makes the music publishers' position stronger than that of the book publishers. It is easier to show a short text has been reproduced than a long one has been absorbed.
This lawsuit arrives as the company prepares its stock market flotation. Litigation of this kind must appear in a prospectus, among the risk factors. The plaintiffs know this. Filing now ensures the case weighs on the valuation, which encourages a quick settlement. It is a classic tactic, and it works.
What it says about the business model
This lawsuit is part of a broader trend we are documenting: training data has become an asset whose holders are seeking to set a price for it.
We have seen it with books bought and destroyed, with expert annotation valued at twenty billion, with a publisher building its own model to avoid licensing its corpus. Music follows the same logic: what was taken for free is paid for afterwards.
The likely direction is a licensing system, as in the deals already struck between music generators and major houses. The lawsuits are not aimed at preventing training; they are aimed at setting its price.
What to take away
This case will probably not go to trial: it will be settled, like the previous one, for an amount that will then serve as a reference for those that follow.
What deserves attention is the mechanism. Each settlement sets a price, each price makes the next lawsuit easier to quantify, and the whole converges towards an economy where training a model includes a cost line that did not exist three years ago.
For the labs, it is one more cost in an already substantial budget. For rights holders, it is the end of a period when their work was used without compensation. And for the GEMA verdict we are still awaiting, it is a reminder that Europe is not alone in having to decide.