Anthropic sued over unlicensed song training

The institution responsible for resolving this dispute is the federal judiciary applying the US Copyright Act through adversarial civil litigation - a mechanism built across a century to adjudicate discrete acts of copying: a printer duplicating a manuscript, a broadcaster airing a recording without a license. It is now being asked to adjudicate something structurally different - the ingestion of tens of thousands of songs into a statistical training process whose outputs are not copies in any traditional sense but weighted probabilities distributed across a neural network. Assess the gap, because the gap is the case.

Sony Music Publishing and Warner Chappell bring the suit as rational-legal bureaucracies in their own right - not as songwriters but as the institutional custodians of songwriters’ rights, aggregators who exist to convert scattered charismatic acts of composition into manageable, licensable, litigable property. This is worth pausing on, because the lawsuit’s moral force depends on an image of the individual composer wronged, while its actual machinery - portfolio management, catalogue valuation, statutory damages calculated per work - is as bureaucratic as anything Anthropic itself operates. Two rationalised systems are colliding, each claiming to speak for something more organic than itself.

The traditional copyright infringement action is designed to answer a narrow question: was this specific work copied, and by whom, and how many times. It answers this through discovery - the production of documents, logs, training manifests - because rational-legal adjudication requires evidence rendered into the form the institution can process. But training a language model is not an act of discrete copying that leaves a discrete trace. It is aggregative and, from outside the company, largely opaque. The contested claim in this case - whether the songs were used to train Claude models at all, and whether “tens of thousands” is the right order of magnitude - is not a disagreement about interpretation but about facts that only Anthropic’s internal records can settle, and Anthropic controls the disclosure of those records. The court, an institution built to weigh evidence, will spend most of its energy simply forcing evidence into existence. This is not a defect of this particular case. It is what happens whenever a bureaucratic procedure engineered for the printing press is redirected at the statistical model.

Consider what actually happens administratively once discovery is complete. The court will not, in practice, issue a philosophical ruling on whether training a model on copyrighted lyrics constitutes fair use in some pure sense. It will produce a settlement architecture - a licensing schedule, a per-track valuation, a compliance framework - because settlement is what rational-legal litigation is operationally built to manufacture when the underlying question is too large for a single verdict and too commercially consequential for either side to risk losing outright. I would direct attention to the analogous machinery already built around sampling disputes and streaming royalties: each time a new technology threatened the publishers’ revenue stream, the eventual resolution was not a clarified principle of authorship but a metering apparatus, a rate card, a clearinghouse. That is the predictable terminus here as well - not a determination of what AI training means for creative labor, but a toll station.

Somewhere in Nashville or Los Angeles a songwriter under contract to one of these publishers will eventually receive a royalty statement with a new line item on it, a few cents attributable to “AI training license,” indistinguishable in format from the mechanical royalty lines that have existed for decades. That songwriter’s actual grievance - the use of a lifetime’s craft as raw material for a system that may eventually generate competing lyrics without them - will have been fully processed by the institution and almost entirely unaddressed by it. The billions in damages sought will be renegotiated down to a manageable royalty regime; Anthropic will treat the settlement as a cost of doing business, the same way every prior generation of media technology treated its own licensing settlements.

The structural prediction, then, is not adjudication but tariff. The courts will not decide what training on copyrighted work means. They will price it.