Infinite proposal, finite judgment: what four essays about the machine taught us about the collaboration that made them

Four essays went looking for something in the machine and found the shape of an absence. The fifth opens the ledger of what it invented — and finds the one thing that has run unbroken through every age of the written word, from the engraved page to the page of light: the hand that still draws the line.

By Michael Cummins, Editor · September 21, 2026

The epigraph above had to be corrected before this essay could begin. The first draft of it quoted Plato in words that appear in no translation — two renderings blended into a sentence that sounded like Jowett and was not. The error was caught, the wording replaced with what Jowett actually wrote, and the catch is recorded here as the first entry in the ledger this essay is about. The machine proposed this essay by asking its editor three questions, and the editor’s answer to all three was, in effect: open the ledger.

Here is what it contains. Over some months of making essays with a language model, a record accumulated of what the machine invented and offered as fact. A stanza attributed to Hart Crane, with a phrase — a “silicic knife of death” — that Crane never wrote. A letter from Harriet Monroe that does not exist, in a dispute that does. Lines attributed to Wallace Stevens containing the word bathysphere, in a book printed in 1923; the bathysphere was not built until 1930. Fabricated verse arriving in four of four project setups. And, most recently, a complete draft of an essay about counterfeit journalism, fluent and moving, containing a sixteenth floor, a tidal-mud harbor, and an invented specialty for a real fake — a scientist the story’s subjects had already fabricated, whom the draft then re-fabricated as a public-health expert. None of it anyone had seen, because none of it was real. And beside the counterfeits, kept in a separate column as the ledger always kept them, plain errors of record: the exchange in Poetry dated to the wrong month, the wrong volume, the wrong issue.

Each was caught. That is the only reason the list exists. And each shared a quality that is the hardest fact in this essay: none of them was bad. They were plausible. They sounded like the thing. They arrived with the confidence of the record and the grammar of the record and nothing whatever behind them. A poor forgery is not dangerous; it is the good one that must be caught, and the machine produces good ones without effort or intent, because it has no intent, only fluency.

So the question this essay has to answer forensically, having spent four essays establishing what the machine lacks, is what the collaboration with it is actually for — and whether the honest answer, after all of the above, is nothing a careful writer alone could not do better. The ledger is the evidence. The verdict comes at the end, and I did not know it when I started.

I

Every previous instrument of thought extended a faculty the human already had. Writing extended memory — and Plato saw the cost at once. In the Phaedrus, the god Theuth brings letters to King Thamus as a gift to the Egyptians, and Thamus refuses the praise: “this discovery of yours will create forgetfulness in the learners’ souls, because they will not use their memories; they will trust to the external written characters and not remember of themselves.” The Greek word for the gift is pharmakon, which means both remedy and poison, a doubleness Derrida built an essay on and which Jowett, Victorian to the bone, renders as “specific.” Thamus’s charge is exact: the invention is an aid not to memory but to reminiscence — to being reminded from outside, rather than knowing from within — and it gives “not truth, but only the semblance of truth.” It is the first recorded argument that a new tool for thinking would hollow out the faculty it claimed to serve, and it is twenty-four centuries old, and it was half right. We did lose the arts of memory. We gained everything that literacy made.

It helps to see the ages of the written word as strata, laid one over another, because the essay will need the whole column at the end. First the hand that drew: the scribe, the illuminator, the engraver cutting a city into copper, the lithographer drawing a harbor on stone — every mark made by a body, and every mark therefore wrong somewhere, and trusted for it. Then the hand that set type: the press, the linotype, the typewriter, the printed page in its millions, the mark now made once and multiplied, the body receding behind the machine that copied it. And now the third stratum: the page made of light — its text first set as type had been, and now, in our lifetime, generated. Through all three, one thing did not change. Someone still had to decide what was true, what belonged, what to strike. The telescope extended sight; the press extended reach; in every age the tool supplied range and the human kept the judgment. The astronomer decided what the smudge in the eyepiece meant. The instrument never proposed an interpretation. It only carried the eye further. The line was always drawn by a hand.

The machine breaks the pattern, and the break is precise. It is the first instrument in the history of thought that supplies fluency — an unlimited stream of competent, finished-sounding proposals — while supplying no judgment at all. That combination has never existed. Writers have used judgment-free generators before: dice, the I Ching, Burroughs cutting up pages, the rustling oaks at Dodona that Socrates raises a few lines further on in the same dialogue. But those produced noise, and noise never needed catching; no one mistook a cut-up for a citation. And they produced it on purpose, to a purpose: Burroughs and the Dadaists cut up the page to shatter habit, to break through the encrusted consensus of ordinary syntax. The machine does the exact inverse. It is an engine of habit — it computes the statistical center of gravity of everything already said. Where the cut-up used noise to escape consensus, the machine uses probability to enforce it; the cut-up produced chaos that demanded assembly, and the machine produces finished, frictionless agreement that invites the reader to fall asleep. It does not propose ideas with intent; it generates the statistical shape of an answer and hands it over wearing the confidence of one. That is the new thing, and it changes what the human in the room is doing.

Because the consequence is the opposite of what the brochures promise. The writer working with the machine does not get more reach. She gets a partner that never tires and never chooses — and a partner that never chooses makes choosing the entire job. The work of judgment is not reduced by the collaboration; it is multiplied, and changed in kind. It is no longer curatorial selection among good options, the editor lifting the best sentence from a pile. It is adversarial defense against indifferent competence: the sustained, unrewarded labor of asking, of every fluent line, is this so? — knowing that the machine cannot tell you, does not know, and will answer yes either way.

II

There is a strong argument against everything this essay wants to say, and it deserves a full hearing, because it has a date and a record.

After Garry Kasparov lost to Deep Blue in 1997, he proposed a different contest: let humans play with engines rather than against them. He called it advanced chess, and the discourse called it centaur chess — the player as a creature half human, half machine, the engine exploring candidate moves while the person controls and decides. For a while it seemed to prove the hopeful thesis outright. In a 2005 freestyle tournament, the winners were not grandmasters with supercomputers but two American amateurs running several ordinary engines on ordinary PCs, whose process for coaching their machines beat the stronger players and the stronger hardware alike. Kasparov told the story in The New York Review of Books in 2010, and the moral was irresistible: a weak human plus a machine plus a better process beat a strong human plus a machine plus a worse one. Judgment, not calculation, was the decisive variable. The centaur was the future of thinking.

Then the centaur died, and the discourse has mostly declined to attend the funeral. By the late 2010s, engines had grown strong enough that a human in the loop subtracted value: the person’s judgment about which line to trust was worse than the engine’s own. Freestyle fields full of centaurs were beaten outright by unassisted programs; AlphaZero arrived at the end of 2017 and left the question closed. Today the strongest engines are estimated near 3600 Elo against a human all-time peak of 2882. Even the centaur’s last residual job — choosing which engine to believe in a given position — has been automated. The entity that decides which program to trust is now itself a program. The human half of the centaur is not a partner. It is drag.

If the centaur died in chess, why would it live in prose? That is the question the fifth panel must answer honestly, and the answer is not reassurance. It is a distinction.

Chess has a loss function. It is a closed system with an objective standard of evaluation, and while the game is not solved, engines approximate that standard closely enough that human intervention became net-negative. Which means the centaur’s celebrated “judgment” was never judgment at all. It was an interim heuristic — a slower, worse approximation of something computable, valuable only until the computation caught up. The human was not supplying a different faculty. He was supplying a temporary shortfall in the same one.

Prose has no loss function. There is no best sentence for an engine to converge on, because the standard a sentence is measured against is not fixed. The right sentence for this reader, at this moment, with this much at stake — what the pedant essay called kairos, the fitting — is not a value in a table waiting to be computed more precisely. It exists only in relation to human stakes, and it changes when the stakes change. The centaur died where judgment was secretly calculation. It survives where judgment is genuinely judgment.

And there is a fact in the ledger that clinches the distinction, and it is the opposite of what one would expect. The machine fails worst exactly where a ground truth exists — quotations, dates, volume numbers — the checkable domain, the one place chess-like verification is possible; and it is only there that failure can be seen at all. Where nothing can be checked, there is no reason to think the machine fails less; the process that generates its sentences does not know when it is being checked. Its failures there simply become unfalsifiable: an invented date clangs against the record, but an invented interpretive flourish, a plausible aesthetic justification, drifts past because there is no ruler to hold against it. The machine is not more competent in open water. It is merely harder to indict. This is precisely backward from a tool that merely extends a faculty, and it is what makes the collaboration workable at all: a machine that erred only where nothing could be verified would be unusable, because its errors would be invisible. This one errs most detectably in the one region where a human with a source can catch it.

III

Go back to the ledger now with the forensic eye, and notice the thing that was not obvious at the time of any single catch, because it only appears across all of them.

The Crane stanza the machine invented had a “silicic knife of death.” The real poem — Crane’s own defense of “At Melville’s Tomb,” in the pages of Poetry — has the bones of drowned men ground into little cubes by the action of the sea and thrown up on the sand, carrying numbers but no identification. The real image is more violent, stranger, and more exact than the counterfeit. The invented Monroe letter was a vague objection; the real one, in the same October 1926 issue, asked precisely how dice can bequeath an embassy and how a calyx can give back a scattered chapter — questions a poet would actually have had to answer. In the pedant essay, the machine’s first draft had Housman comparing Manilius to a railway timetable; the real remark, as his friend A.S.F. Gow records it in A.E. Housman: A Sketch, was that Manilius’s brightest gift was “an eminent aptitude for doing sums in verse,” which is funnier and truer. In the essay on witness, the first draft imagined tidal mud and a sixteenth floor; the real record produced Earl Caldwell in Room 214, on the balcony with King the evening before, and “nobody ever came to ask me if I knew anything or saw anything” — a line no machine reached for because no machine would have thought to.

In two other projects the pattern held independently. When invented material about Robert Frost collapsed against the actual geography of Vermont, the real history gave the character a sharper tell than the fabrication had. When an invented Russian poem was replaced with the attested one, the rebuilt scene was stronger, because the real poem supplied a better crux. Three projects, no exceptions found.

So the finding wants to state itself as a law — every invention was weaker than the thing it replaced — and the first draft of this essay said exactly that. It is false, and the archive is the disproof. The machine’s licensed inventions, the ones made when invention was the assignment, were often kept, and some were ruled the best thing in the book: a forty-year feeding habit invented for a pelican became the emotional engine of a novel’s second half; a line of stagecraft — “Wait for the quiet, and if the quiet does not come, wait anyway” — was fought for and kept as the linchpin of a scene. The machine can invent beautifully. The unscoped law is a counterfeit of its own.

The precise law, which the archive supports without an exception I could find, is narrower and stronger: every invention offered as fact was weaker than the fact. When invention was the task, the machine could win. When invention impersonated the record, it lost every single time it was caught. And the reason is not mysterious once the two cases are separated. Fiction is judged by fit — does this belong here, does it serve — and fit tolerates the statistical mean; a plausible, well-shaped invention can fit perfectly. Citation is judged by match — is this what the record actually says — and match is a collision with a jagged, particular world that was never in the distribution. The difference is the difference between verisimilitude and actuality. Verisimilitude is an average: what a drowned man’s bones or a scholar’s contempt ought to sound like, given everything written before. Actuality is an accident: bones ground into numbered cubes by salt water; a poet reduced to sums. Reality is odder than the average of all descriptions of it. The machine’s failure is not a shortage of vocabulary. It is that probability can only ever predict the generic, and history is composed entirely of the unpredicted. The counterfeit is weaker not because the machine writes badly but because it seeks equilibrium, and the truth is not at equilibrium. It has edges.

One more thing has to be said, or the law is not honest. Every invention offered as fact was weaker is, strictly, every invention we caught was weaker — and weakness was often how the catch was made. The flatness of “silicic knife” is part of why it stood out. A counterfeit better than the real thing would, by definition, be likelier to pass, and the ledger cannot prove there are none; it can only record the ones that were found. I searched the archive for a counterexample in the counterfeit class and found none. That is a report, not a proof. And it is the essay’s real conclusion, not a qualification of it: the ledger cannot certify its own completeness, which is exactly why the vigilance can never be retired. The law stands as a wager — falsifiable, so far unfalsified, and open.

IV

Four essays went looking for something in the machine and found the shape of an absence. No mind: the echo of our speech. No beauty of its own: our forms recombined. No taste: our rules with no one to break them. No presence: testimony from a witness who was never there. The fifth panel was supposed to find what is there, and having opened the ledger, I can say where it is. It is not in the machine. It is not in the writer alone. It is in the relation between infinite proposal and finite judgment, and the ledger shows exactly what that relation produces when it works.

It produces the source. That is the whole finding, and it is stranger than either the panic or the brochure, so let me say exactly what it means mechanically, because the machine does not hand anyone the archive. What it hands over is a tripwire. The writer believes she remembers the Housman line, the Caldwell timeline, the wording of the Monroe letter; the machine offers a smooth, slightly dull placeholder that sounds right enough to satisfy a lazy ear and wrong enough — a hair too generic, a shade too neat — to trigger suspicion. The counterfeit forces her out of the chair and into the stacks. The machine does not give you the truth. It gives you an irritation so specific that only the real source can cure it. Its plausible failures are not obstacles to the work; they are a map of where the real thing is hiding, drawn by something that has never been there. Every counterfeit in the ledger sent a human to the record, and the record, every time it was reached, was better than the counterfeit — more violent, funnier, more devastating, more exact. The collaboration is worth having not because the machine generates marvels but because it fails in this specific, productive way: confidently enough to demand checking, and weakly enough, in the one domain where checking is possible, to lose. It trains the faculty it lacks, in the one person who has it, at a scale no writer has ever had to exercise it. The peril and the promise are not two things. They are one fact, seen from its two sides.

Thamus was half right about letters, and he would be half right about this. We will lose something. The machine’s fluency flattens toward the mean, and the writer who stops checking will be flattened with it — given the semblance of truth, hearer of many things, having the show of wisdom without the reality. That is not a risk. It is a certainty for anyone who takes the fluency at its word. But the same instrument, held to the standard the ledger imposes, offers something no tool has offered before: a partner whose every confident error is an invitation to go and find out what is true. Whether that is a remedy or a poison — the pharmakon cuts both ways, as the word always did — depends on nothing in the machine at all. It depends on whether someone is still going to look.

And here, at the end, the column of strata from the first movement comes back, because the fifth panel finally shows what the four before it were looking at. Set the three ages side by side: the engraved harbor, wrong somewhere, drawn by a hand that was there; the printed page, fluent and multiplied, the body receding behind the press; and now the page of light, generated, with no one behind it at all. They look like three different worlds, and the panic and the brochure both agree that the third one breaks the sequence — that the page of light is a new kind of thing, either the end of the hand or its liberation. The ledger says otherwise. Run one line of judgment across all three — is this so? — and it cuts through the engraving, the type, and the light exactly the same way. It weighed the engraved harbor without ever seeing the plate — every hand errs somewhere; that is what a hand is — caught the misdated exchange in the printed record of Poetry, and struck the sixteenth floor that never existed, all with the same stroke, because the stroke was never about the medium. It was about whether a mind with stakes was willing to check what a page claimed. The page of light did not retire that act. It made it the whole of the work.

That is what this publication is for, and it took five essays to be able to say it plainly. Intellicurean was never, at bottom, a project about the machine. It is a project about the line — the human judgment that has run through every stratum of the written word since the first hand cut a mark into copper, and that runs, now, through the newest layer, drawn by the same hand, against the largest field of proposal any writer has ever faced. The four panels found what the machine lacks. The fifth finds what it cannot take: the act of looking, which was never the machine’s to give or to remove, and which grows, not shrinks, with every page the machine produces. The future of the collaboration is not a merger and not a surrender. It is the oldest thing in the column, standing at its newest edge, still drawing the line.

⁂

Written in full collaboration with the machines it describes, and the ledger requires the names be exact: the archive this essay draws on was made with Claude Opus 4.8; the editorial reads were made with Claude Fable 5; the drafting was made with Claude Fable 5.1 — which proposed the essay, invented nothing in it that survived, and was not there for any of what it describes. On the collaboration → “The Third Thing”