Anthropic sued over unlicensed song training

This lawsuit benefits perhaps tens of thousands of songwriters and composers, whose work was allegedly taken to train a system worth billions, by the prospect of compensation and by the deterrent effect a large verdict would have on future takings. It harms Anthropic’s investors and, more diffusely, the millions of people who use Claude for work they could not otherwise afford, by the prospect of billions in damages, a slower pipeline of AI development, and a legal cloud over every model trained since. The arithmetic is uncomfortable, but the arithmetic is the argument, and it does not resolve as cleanly as either side’s lawyers would like.

Let us count properly. Sony Music Publishing and Warner Chappell are not songwriters. They are administrators of songwriters’ rights, and they will take their customary cut of any settlement before a single dollar reaches the composer who wrote a hook in 1987 and has been living off quarterly statements ever since. This matters to the calculus because the pain being redressed and the party collecting the redress are not the same party in the same proportion. A session writer whose three-minute song sits somewhere in Anthropic’s training data suffers a real diminishment: not merely lost licensing fees, but the particular indignity of discovering that one’s labour was folded into a machine without so much as a request. That pain is diffuse across tens of thousands of catalogues and concentrated, in practice, into two publishing houses large enough to litigate. Each songwriter counts for one. But two publishers, not thirty thousand songwriters, will decide what the case is worth and what a settlement buys.

Now weigh the other side of the ledger. Anthropic’s defence, stripped of legal jargon, is that training a model on copyrighted material to learn patterns of language is not the same act as reproducing that material for consumption - a distinction with real utilitarian weight, since the harm of piracy is substitution, and a model that has merely absorbed statistical regularities from a song is not obviously substituting for that song in the market. I take this argument seriously, because I am suspicious of any moralist who treats “using copyrighted work” as self-evidently wrong regardless of what use is made. If Claude does not reproduce recognisable lyrics, the songwriter has lost nothing that a licence would have protected in the first place. The dispute over “tens of thousands” of songs, and over whether they specifically trained the Claude models in question, is therefore not a rounding error - it is the entire question of whether any concrete pain occurred at all, as opposed to a violation of principle with no measurable victim.

Here is where the calculus turns unfashionable. Suppose the pain is real but small per song, and the number of songs is large. Multiplied across tens of thousands of works, a small individual harm becomes a substantial aggregate one - this is simply arithmetic, and I will not pretend otherwise to flatter the AI industry. But suppose also that resolving this dispute through litigation, rather than through an upfront licensing market, produces years of uncertainty that deters every AI firm from documenting what it trains on, which in turn makes future violations harder to detect and future settlements harder to value. A regime of retrospective lawsuits teaches companies to obscure their training data, not to license it honestly. That is the perverse incentive a rational legislator must attend to: punishment applied after the fact, with no clear prior tariff, produces concealment rather than compliance.

The reform implication is not sympathy for Anthropic, nor reflexive solidarity with songwriters, but a demand for a standing licensing mechanism - a compulsory rate, set in advance, the way radio royalties were settled a century ago rather than litigated song by song. Such a scheme would compensate the composer whose hook sits inside a model without turning every training run into a decade of discovery motions benefiting mainly the publishers and the lawyers. Until legislators build that mechanism, every verdict in cases like this one is not justice calculated but justice improvised, and improvised justice, however large the number attached to it, rarely finds its way back to the person who wrote the song.