Sparks: Anthropic sued over alleged theft of ‘tens of thousands’ of songs

They call it training, but what we now call training was, before this analysis, merely the unacknowledged consumption of another man's labour - and a practice that cannot name what it takes cannot be trusted to measure it.

If the song could not have been made without the songwriter, and the machine could not have been made without the song, then the question is not who owns the machine but who has been paid for the ladder he climbed.

Whatever the lawyers decide, the cash-value question is simpler: on Tuesday morning, does the songwriter eat, and what difference does any theory of machine learning make to his breakfast?

Forget the arguments about art and progress; here is who has leverage - publishers holding statutory damages, a startup holding valuation - and whoever settles first pays least, which is the whole of the strategy.

No tribunal of public sentiment will restrain an industry that profits from what it takes; only a licensing mechanism that makes payment cheaper than litigation will produce the conduct virtue never has.

The company's statement spoke of innovation, fair use, and the future of human creativity; nowhere in three paragraphs did the word songwriter appear, and everyone read it anyway.

Everyone in this dispute treats the phrase 'training data' as a natural category like air or water, and that common sense - data as raw material free for the taking - is precisely the assumption the lawsuit exists to interrupt.

The boardroom had agreed, with impeccable courtesy, that all creative work ultimately belongs to humanity, a proposition advanced chiefly by gentlemen who had never created anything and owned a great deal.

In this best of all possible worlds, the company took the songs without payment, called the taking progress, and the songwriters were invited to be grateful for the acceleration of culture at their expense.

They built the machine on the songs, denied the songs, and called the denial innovation - the idea will outlive the lawsuit, but the burning tells you who held the matches.

'Tens of thousands of songs were used to train the model.' Note the grammar: the songs used themselves, the model trained itself, and somewhere between the passive verb and the missing agent, every songwriter vanished from their own sentence.

In every market I have walked, from Delhi to Fez, the merchant who takes goods without weighing coin for them is called a thief, and I have yet to visit the city where the name changes because the goods were songs.

Things that are hateful: a machine that sings beautifully, a ledger with no entry where the songs should be, and an apology that arrives only after the damages are counted.