Machines are still only in their trial period. One day, perhaps, an electrophonic kaleidoscope will compose musical figures in their thousands — will invent whole series of rhythms and melodies. Man will have machines that create just as he will have machines for precision thinking — His role will be reduced to choosing. Combination factories.
The most intense feeling is only one combination among others.
— Paul Valéry, Cahiers / Notebooks 3
The standard complaint about large language models is that they will make writers sound alike, and perhaps eventually think alike. But Valéry’s prescient suggestion makes me wonder about another possibility. What if these machines do not merely automate an existing kind of writing? What if they also call forth a somewhat different skill in writers?
We ordinarily imagine writing in painterly terms. The writer confronts a blank surface and begins putting things onto it. A word, then a sentence, then a paragraph. The work accumulates. Even when we know perfectly well that revision will follow, our basic picture of authorship remains additive: there was nothing here, and now the writer has made something out of his inner resources and aesthetic vision.
An LLM makes possible a more sculptural way of working. Instead of confronting a blank page, the writer can begin with an excess of material. The machine can produce the boulder of text. The writer’s task is to decide what, if anything, is inside it. That sounds close to Valéry’s prediction that our role will be “reduced to choosing.” But choosing may be doing a lot of work there. A sculptor also “chooses.” This part stays; that part goes. This line needs to emerge; that mass is obscuring it. The finished form depends not on the ability to manufacture marble but on seeing possibilities in it that someone else would not see. Two sculptors can begin with approximately the same block and end with objects bearing no resemblance to one another.
Give two people the same serviceable paragraph from an LLM. One may read it and think it’s fine as is. The other notices immediately that the opening sentence is unnecessary, the metaphor in the middle is the only interesting thing there, the conclusion merely repeats the premise, and one odd phrase suggests a better direction for the whole piece. This would shift some of the emphasis in writing from generation toward discrimination. An LLM can give you ten openings before you have finished thinking about the first. It can propose twenty metaphors, five structures, three counterarguments, a dozen ways of phrasing the sentence you are stuck on. Under those conditions, producing another possibility is nearly worthless, but knowing which possibility is promising becomes more valuable. That sounds less like the traditional ideal of the writer and more like the work of an editor.
I haven’t experimented with this yet. I don’t know how it might go in practice. It probably would not go one single way for everybody. But this seems more interesting than treating AI-assisted writing solely as a question of whether the machine has “done the writing.” That formulation assumes we already know where the important part of writing resides. Well, maybe. Or maybe abundance changes where the difficulty lies. The blank page tests one set of capacities; the overflowing page may test another.