Who Owns a Sentence

In early 2020, a lawyer and coder named Damien Riehl explained from a stage how he and a musician named Noah Rubin set out to end a certain kind of lawsuit by committing it to absurdity. They had written a small program that generated melodies; every melody, or close enough. Their program produced every possible run of eight notes across twelve beats, some sixty-eight billion of them. They then released the whole catalogue into the public domain.^[1] The point was an argument of exhaustion, music beside it; no one will ever play the hard drive. There are only so many ways to arrange a few notes into a line a person can hum. The number is enormous, but it is finite, and every one of those lines already existed, as a mathematical fact, before any songwriter was born to write it. A melody, Riehl was saying, is closer to found than invented, found the way the number seventeen is found, and nobody owns seventeen.

What is true of melodies is, with complications, true of sentences. The match is loose; language is vastly larger than an eight-note run, and a paragraph has more room to be singular than a chord progression does. But the direction is the same, and it points somewhere uncomfortable for our instincts about ownership. Almost everything you will ever write is recombination. The words are inherited; the grammar is inherited; the moves (the turn, the reversal, the concession, the close on a hard image) were all worn smooth by other hands long before you reached for them. The metaphor that feels like yours has a genealogy. The argument that feels new is, nine times out of ten, an old argument wearing this decade’s clothes. We are each standing on a mountain of inherited competence we did not build, and writing is mostly the act of rearranging what is already on that mountain.

This should be less of a scandal than it feels, because the idea that it is a scandal is surprisingly young. The figure of the wholly original genius, the creator who spins the new thing out of private inspiration, is largely an invention of the Romantic era, and it was always more flattering than accurate. Shakespeare took his plots wherever he found them: Holinshed’s chronicles, Plutarch’s lives, a Danish prince. The borrowing is half the genius, and he stands taller for it. What he did with Holinshed is the whole point, and Holinshed is the one forgotten. T. S. Eliot put it in a sentence that has itself been borrowed ever since: immature poets imitate, mature poets steal.^[2] Bach reworked. The blues reworked. The artist and creator rework. For most of written history, the commonplace book was ordinary equipment for a thinking life, a tool a serious mind kept close. Originality, in the strong sense the copyright lawsuit assumes, may be the rarest thing there is, and may be a mirage entirely. What is real is a particular person doing a particular rearranging.

Jorge Luis Borges wrote the cleanest demonstration of this in 1939, in a story about a man who set out to write Don Quixote. He set out to compose it again, word for word, the same book Cervantes wrote, arrived at independently by a twentieth-century Frenchman.^[3] Borges quotes a passage from Cervantes and the identical passage from his fictional author, Pierre Menard, and argues, with a straight face, that Menard’s is the richer of the two, because the same words, written three hundred years later by a different man in a different world, mean differently. The phrase that was ordinary in Cervantes’s mouth is, in Menard’s, a deliberate archaism. The text is identical. The achievement is not. Borges meant it as a joke and a profundity: the value lived in who set the words down, at what cost, and standing where. A later chapter will make the same point with a painting instead of a sentence: a single unchanged canvas counted a masterpiece or a worthless fake depending only on whose hand was believed to be behind it. A forgery and a sentence have the same secret. The part that mattered always lay outside the copy.

Now turn to the thing that has writers in open revolt: machines trained on everything, generating sentences in your cadence, images in your style, and the claim hardening into lawsuits across two continents: that’s mine, you stole it. Two true things are happening here, and the whole argument depends on pulling them apart instead of letting the one you can answer swallow the one you cannot.

The first is the claim instincts reach for first. Your sentences were recombination, like everyone’s; your style is a probability distribution over moves you learned from others; the model that read ten thousand writers and the novelist who read a thousand are, at the level of the sentence, doing a recognizable version of the same thing. We have a word for the one who learns the moves by watching another work until the moves become his own: apprentice. You copied me proves less than it wants to, because copying in that loose sense is what every writer who ever got good always did.

The second claim is the one the recombination argument is built to walk past. The gap between the student and the model is not only scale, though scale is where it opens. The novelist who reads a thousand books feeds the world those books came from: more readers, a few more sales, sometimes a career that points back toward the writers who formed her. A company that ingests a million books to build a machine that competes with their authors sends nothing back up the line. The corpus was a working input to a product sold for profit, the input was taken without consent or payment, and the gain it created is kept by the party that took it. That is extraction, and it stays extraction whether or not a sentence can be owned. Whether language is property has no bearing on whether a worker was paid. The New York Times, the novelists in the Authors Guild suit, the illustrators who watch their own hand come back out of a prompt box are owed a reply in the currency they are asking in, and the courts may yet decide the currency is money. What that reply should be built from, a per-use market that pays each contributor when their work moves an output, or a class-wide settlement that splits a lump sum across everyone scraped, is genuinely unsettled, and people who agree the debt is real divide hard over which one to want. This book argues where the value belongs, not which machinery should deliver it. That the machinery is still unbuilt is the urgent, unfinished business the argument cannot itself close.^[4]

Underneath the lawsuit sits a second grief, and the legal vocabulary keeps mistranslating it because money cannot reach it. The thing that was ever fully yours lived off the page, which is why it survived everything the model could do to the page: the stake. That a particular, findable, answerable person stood behind this claim and would pay if it failed. The machine reproduces your style and stops at your stakes, the same wall the forger met between the brushwork and the provenance. Your words were copied into a place where no one is answerable for them, absorbed into an averaged, sourceless output that wears your cadence while shedding your exposure, until the link between an utterance and the person who meant it is cut at industrial scale. What breaks there is attribution, the chain that runs back to an answerable source, and it sits outside copyright’s reach and outside compensation’s too. This is a separate loss from the unpaid wage. A writer can be paid in full and still suffer it; a writer can be credited by name and still suffer it. The check answers the theft of labor and leaves this one standing, because no check was ever the thing that held a sentence to the person who would answer for it.

All of this leaves the proprietary book standing. A serious printed work is its author’s, and the claim rests on something other than owning the sentences. It rests on being the one who spent the years selecting, synthesizing, and staking a name on this account of the subject, this one above all others. The recombination is curated, tested, and answered for by a specific person, and that is a real and legitimate claim. The book holds its weight; copyright simply reaches for the value at the wrong layer. The value the author actually created, and may rightly charge for, is the accountability they bring to the synthesis, which runs deeper than the copyability of its sentences.

“Who owns this sentence” is the wrong question even when the grievance behind it is right. Ideas were never owned; truth least of all. No one held a property right in the boiling point of water or the structure of a tragedy, and a civilization that tried to would have strangled itself within its first generation. What a healthy culture tracks is who will answer for a true thing now. Who first uttered it usually vanishes and usually doesn’t matter; what counts is whose name is on the study, whose reputation rides on the claim, who can be found and asked how do you know. Attestation is the relation that carries the weight: a bond between a claim and a person, drawn through reputation rather than around a string of words.

The machines have made strings of words free, as they have made melodies free, as printing once made copies free. Each time, the people who had priced their living in the copyable thing felt the floor give. Whether they were paid as it gave was never settled by the machine; it was settled by who held the leverage and what the law chose to protect, and it has come out one way in some rooms and the opposite in others. The license, the royalty, the collective bargain are live in this round too, and scraped writers may well be paid. What no settlement reaches is the loss that ran the other way, into the part that survives copying, where the value sat all along: the answerable person behind the words, who carries a price no licensing rate can name.

So put the sentence down; the asset always lay elsewhere. The question worth asking of any made thing in the age of free expression reaches past whether the words are original. Almost no words are, and the machine will finish off whatever pretense still survives. The question is the older one, the one a melody on a hard drive can never raise and a person always can: not who wrote this, but who will stand behind it when it matters, and what it will cost them if they are wrong.

Notes (4)

1. 2020; released under a Creative Commons Zero license as the “All the Music” project. Techdirt; Kottke, Feb. 2020. ↩︎

2. T. S. Eliot, “Philip Massinger,” in The Sacred Wood: Essays on Poetry and Criticism (1920). The full passage: “Immature poets imitate; mature poets steal; bad poets deface what they take, and good poets make it into something better, or at least something different. The good poet welds his theft into a whole of feeling which is unique, utterly different from that from which it was torn.” ↩︎

3. Jorge Luis Borges, “Pierre Menard, Author of the Quixote” (1939), whose protagonist recomposes Don Quixote word for word without copying it. First published in the Argentine journal Sur; collected in Ficciones (1944). ↩︎

4. As of mid-2026, whether training on copyrighted material constitutes infringement remains unresolved in US courts. The New York Times sued OpenAI and Microsoft in December 2023, alleging that wholesale ingestion of its articles exceeded fair use; the case survived a motion to dismiss in February 2025 and is proceeding to discovery. Class actions from authors including George R. R. Martin, Jodi Picoult, and others through the Authors Guild, and Getty Images’ suits against Stability AI in both US and UK courts, present related but distinct theories. The central unresolved question is whether large-scale statistical training constitutes “transformative use” as courts have found for some search-engine indexes, or is closer to the systematic reproduction the Times alleges. The EU AI Act (in force 2024) requires disclosure of training data but does not resolve underlying copyright questions. Separately from the infringement question, consent and compensation frameworks for training data are under development at both legislative and industry-standards levels; several major publishers have reached licensing agreements with AI companies while litigation proceeds. ↩︎