Microsoft Copilot rarely copies full sentences, firm claims
The story celebrates that Copilot was built - the feat of a machine that can read the accumulated sentences of the world and answer in kind. Microsoft’s own filings now frame this achievement in the negative: the system rarely reproduces full sentences from news articles and books. But a made thing does not stop where its maker’s assurances stop; it goes on acting in a world no legal brief contains. The question the filing skips is the only one that lasts: who is answerable for what this does after release, and what did Microsoft fail to imagine when it built a machine to read everything and repeat some of it back?
Notice what “rarely” is doing in that sentence. It is not a denial; it is an admission wearing the clothes of a defense. Microsoft is not claiming Copilot never reproduces the work of The New York Times or of book authors - it is claiming the reproduction is infrequent enough not to matter. This is the maker speaking in the voice of statistics rather than the voice of consequence, and it is worth asking why. A press release measures triumph in what the system can do. A legal filing measures liability in what the system can be shown, on average, not to do. Between those two modes of speech sits the entire distance between the moment of creation and the moment of reckoning.
The New York Times and the book authors are not quarrelling with Copilot’s competence; they are quarrelling with its appetite. The claim of substitution - that a reader might get from the machine what they once paid the author for - does not require constant reproduction to be true. It requires only that the machine be capable of it, at moments neither predictable nor rare enough to be dismissed as noise. A machine that reproduces a paragraph of a novel once in ten thousand queries has still, in that instance, done the thing the author’s contract with her publisher exists to prevent. Frequency is a defense against embarrassment. It is not a defense against harm.
What Microsoft failed to imagine - or, more precisely, imagined and set aside - is that a system built to answer in the register of everything it has read cannot be neatly partitioned into “reproduction” and “generation” after the fact. The company that trained Copilot on the corpus of human writing made a wager: that the system’s outputs would blur into something legally distinct from their sources often enough that the exceptions could be treated as exceptions. That wager was made at the point of training, months or years before any court asked to see the evidence. The responsibility for what the system does now was incurred then, not at the moment the litigation began. This is the quiet handoff that always follows a launch: the builders spoke of capability, and now, when the capability turns out to have edges, the argument becomes about frequency, about rarity, about how seldom the seams show.
Picture the single sentence lifted whole from a book, sitting inside a chatbot’s answer to someone who never bought that book and never intended to. The author of that sentence spent, perhaps, a year earning the right to it - revising, cutting, testing it against an editor’s doubt. The machine offers it up in half a second, unbilled, unattributed, indistinguishable in the moment from the machine’s own words. That is not a statistical rarity to the author whose sentence it was. It is the entire transaction of authorship, undone in an instant by a system whose maker is now arguing, in the register of a courtroom, that this sort of thing hardly ever happens.
The wisdom to be accountable for a system like this would have required Microsoft to build not just a machine that reads everything, but a mechanism answerable for everything it might say back - before the filings, before the lawsuits, before “rarely” became the load-bearing word in the company’s defense. The capability came first. The accountability is still being negotiated, one contested sentence at a time, in rooms far from the lab where the thing was made.