Chapter Six
The Veto
In November 1924, Maxwell Perkins of Charles Scribner’s Sons sat down to write to a twenty-eight-year-old author whose new manuscript had just crossed his desk. The letter opens with the praise the book deserved: “I think the novel is a wonder… It has vitality to an extraordinary degree, and glamour, and a great deal of underlying thought of unusual quality.” Perkins meant every word; he had fought his own house’s senior editors to publish this writer in the first place. Then, having meant it, he did his job. “Among a set of characters marvelously palpable and vital,” he wrote, “I would know Tom Buchanan if I met him on the street and would avoid him—Gatsby is somewhat vague. The reader’s eyes can never quite focus upon him, his outlines are dim.”[1]
The manuscript was The Great Gatsby. It was finished, it was publishable exactly as it stood, and the figure the entire novel hung on was, in the judgment of its editor, out of focus. Perkins did not soften the point into a question or bury it among the compliments. He developed it for paragraphs, and he prescribed no fix; he proposed no replacement lines of his own. He only asked questions. Couldn’t Gatsby be described as distinctly as the others were? Couldn’t his strange past be hinted at along the way rather than withheld until the end? The whole letter pointed, without dictating, at the one place the book had not yet become the thing it was reaching for. Fitzgerald, who could have refused, instead spent the winter rebuilding Gatsby from the inside, redistributing the revelations of his past, sharpening his physical presence, manufacturing the famous smile that “believed in you as you would like to believe in yourself,” and wrote to Perkins that he had brought Gatsby to life at last, and that he could now write whatever came next with confidence. The novel published in April 1925 is the one the letter demanded. The one that arrived in November 1924, competent and complete and almost, is the one nobody would have remembered.
Everything the previous chapter called a stake, Perkins was holding at once: a reputation staked before the small audience of people in publishing who could judge an editor’s eye; a falsifiable commitment (champion this writer, against his own house’s resistance, a wager the world would score); a standard, visible across a career of such letters, that any given book either met or did not. And then, the part that distinguishes him from every thoughtful person who has ever read a flawed manuscript: he was positioned to act on the judgment, and he acted, at the moment of maximum friction, when the book was done and the schedule was waiting. The judgment, the stake, and the power to refuse, joined in one person at the decisive moment. That conjunction is what this book calls the veto, and everything durable in our culture’s quality has depended on it.
The veto is not quality control. Quality control checks work against a specification: word count, fact pattern, house style, test suite. It is honorable, necessary work, and it is in principle automatable, because a specification is exactly the kind of thing a machine can check. The Gatsby manuscript passed every specification; Perkins says so in the first line. The novel was a wonder. The veto operates above the specification. It asks the question every checklist leaves out: given what this particular thing is trying to be, is it there yet? The answer lives in the gap between the work and a standard that exists only in someone’s trained judgment, beyond anything the spec can compute, and the answer only matters if the someone is willing to pay the cost of giving it.[2]
Every domain that has maintained quality over time has built a seat for this person.
Thelma Schoonmaker has edited nearly every Martin Scorsese feature since Raging Bull in 1980: forty-five years, three Oscars, as durable a director-editor partnership as American film has produced.[3] The popular understanding of her job is assembly. The actual job, as both of them have described it across decades of interviews, is closer to adjudication: she is the person inside the process empowered to tell one of the most celebrated directors alive that a scene (beautifully shot, expensively staged, technically flawless) is not working, and to cut it. Scorsese, by any measure one of the most accomplished living masters of film craft, keeps a person whose role is to overrule him, because he understands that the maker’s relationship to his own footage is compromised by the cost of having shot it. What makes her veto function is the same triple Perkins had: a judgment built across ten thousand hours in the cutting room, a reputation fused to the films (her name is on every one; the failures would be hers too), and a structural arrangement in which her no is real. Strip any leg and the seat collapses: judgment without power is a consultant; power without judgment is a hazard; both without stakes is a bureaucrat.
Peer review, in the fields where it still functions, is the same seat built for knowledge instead of art. The reviewer who writes “the methodology does not support the conclusion” is exercising a Perkins veto on a manuscript that may represent years of someone’s funded life. The reviewer is anonymous to the author, which looks stakeless until you notice where the stakes actually sit: with the editor, who knows the reviewer’s name and keeps, over a career, a private ledger of whose judgments held up. A reviewer who waves through flawed work, or kills sound work out of rivalry, spends down an account that journal editors track with the unsentimental memory of bankers. The anonymity shields the reviewer from the author’s retaliation while leaving him fully exposed to the editor’s bookkeeping. When people say peer review is broken (and in many places it is), the breakage is almost always describable in this chapter’s terms: the volume of submissions outran the supply of staked judgment, and the seats got filled with reviewers whose carelessness costs them nothing that anyone records. Volume is not the only corrosion. Reviewers settle scores, miss outright fraud, nod through fashionable error, defer to a famous name; the editor’s private ledger that is supposed to catch them is itself only as honest as the editor keeping it. But the structural claim survives every one of these failures: peer review works to exactly the degree that someone’s standing actually rides on the judgment, and rots at exactly the rate that stops being true.
The cost of this position is most visible when the words themselves are technically flawless. A professor is asked to write a letter of recommendation for a student she barely taught. She is busy; the deadline is tight; she prompts a model with the student’s transcript and a few details, and the output is excellent: specific-sounding, warm, persuasive, better than what she would have managed between meetings.
But the letter is slop, and not because a machine wrote it. It is slop because no one stands behind the claim it makes. A recommendation was never valued for its prose; its whole worth is that a known person puts her own standing on the line to vouch for another, so the reader can lean on her judgment instead of guessing. When she has no judgment to give, because she never knew the work, the warm sentences certify nothing. The reader is trusting an assessment no one actually made. The “No” she might have said, that she cannot in good conscience vouch for a student she did not know, was the one thing the task was really asking for, and it is exactly what got automated away. The dignity of the recommendation was never in its eloquence; it lives in her answerability for a judgment she actually made. Here there was no such judgment to begin with, so the warm sentences certify an assessment that was never formed. The fault is not that a machine found the words; had she known the student cold, she could have let a model phrase her verdict and stood behind every line of it. The fault is that nothing underneath the words was ever staked, and no fluency supplies what was never there.
The point is not about the machine, and here is the test that proves it. Picture the same professor recommending a student she mentored for three years, whose work she knows cold, who then, late and tired, asks a model to help her find clean language for an assessment she has already made. That letter escapes slop, even if some of its sentences came from the same place the first one’s did. The judgment is real; the vouching is staked; a person stands behind every line and would feel it as her own failure if the student flamed out. The absence is the thing, and the tool is incidental to it. Slop is what you get when no one made the judgment, and a machine cannot make one; what it equally cannot do is unmake the judgment of someone who did.
Now set against all of this the way machine-age content is actually produced, and locate the veto in it. A model drafts; perhaps a second model scores the draft; a human, if a human appears at all, performs what the industry calls review, which at production volumes means seconds of glance per item; the content ships. Where in this chain is the Perkins question even askable? The scoring model checks the spec: it is quality control, useful, constitutionally incapable of asking whether the spec is the right spec for this particular thing, because “this particular thing trying to be something” lies outside the concepts available to it. The human reviewer at volume is triaging in place of judging, and the economics of her position have been arranged precisely so that sending work back is the one thing she cannot afford to do.
The industry often confuses this triage with “content strategy.” Content strategy is the optimization of output toward a goal: engagement, conversion, reach. It is a set of “Yeses” stacked on top of each other. The veto is the power of the “No” that makes the “Yes” valuable. It is the friction that ensures the signal is worth the noise. Strategy asks “how can we get this to more people?”; the veto asks “should this exist at all?” When you remove the cost of the refusal, you remove the value of the acceptance. You are left with a system that can produce everything and stand behind nothing.
“The feedback loop will fix it” is the industry’s standing answer, and the training process itself falls short of a veto too. Reinforcement from human feedback tunes a system toward outputs that raters prefer on average, which is the statistical opposite of the Perkins letter: a veto is one staked person’s judgment about one particular work against the average preference. Averaging is how you manufacture adequacy. It is structurally incapable of manufacturing the thing Perkins was protecting, which begins exactly where adequacy ends.
So when production is automated and the veto seat is left unfilled, the output is slop by construction, slop as a standing condition rather than an occasional lapse, content from a chain in which no one was positioned to be embarrassed, regardless of how good the model is, and it gets more dangerous as the model gets better, because better models clear higher specifications, and the spec was never the question.
But suppose a machine does not only draft. Suppose one finishes. Suppose the model produces, on the first pass, the perfect Gatsby: the very book Perkins’s letter was reaching for, arrived at without the letter. Doesn’t that make the veto a piece of obsolete friction, a stage the machine has learned to skip?
It would, if the veto were only correction. But correction is the smaller of the two things a veto does. The larger is recognition. The veto is the faculty that knows a perfect Gatsby when it sees one. And the same model that could generate the perfect Gatsby on one pass will generate ten thousand confident near-misses on the next ten thousand, with nothing in the output to announce which is which. A masterpiece and an accomplished fake arrive in identical packaging, as the forged Vermeer and the real one did; the difference is never on the surface. So even in the imagined case, someone has to be able to tell, and telling, measuring a particular work against a standard held in trained judgment, is the veto. Remove it and what goes with it is the only thing that could have known the machine got it right, the bottleneck turning out to have been the faculty of recognition all along.
And there is a second thing the perfect Gatsby still lacks, which no quality of the text can supply: someone answerable for it. The reader who trusts Gatsby is trusting a chain that runs back through Fitzgerald’s name and Scribner’s masthead and Perkins’s judgment to people who staked themselves on it. A flawless Gatsby generated by no one and stood behind by no one is not a fraud; chapter four gave the authorless wonder its due. But it arrives uninsured and, worse, unfindable: one perfect book in a sea of ten thousand confident near-misses wearing the same cover, with no one positioned to tell you, and your evenings are too finite to know which is which. What the chain was selling went beyond the novel itself. It was the prior promise that this one, out of all the manuscripts that crossed the desk that year, deserved the price of your attention: a promise someone had to be capable of making, and of being publicly wrong about. The veto is the human position in the chain: the one who can recognize, and the one who will answer. A better model empties neither seat.
The posture that follows from this is the reverse of refusal. If the machine can out-draft you (and on a widening range of work it now can), then competing with it at drafting is a fool’s contest, and spurning it for purity’s sake is mere sentiment. The intelligent move is to take everything it offers and occupy the two seats it cannot fill: be the one who recognizes when a draft has reached the thing it was straining to become, and the one who answers when it has not. Used that way the machine is the most powerful instrument for quality ever put in a maker’s hands, capable of lifting a competent draft toward what it was reaching for faster than any editor in history, a Perkins available at three in the morning and never once tired. The single thing it cannot do is hold the stake. Consulting the most capable mind in the room was always sound; the error comes from letting that mind also be the one who would be embarrassed, which it can never be. Borrow all the intelligence you can reach, then sign your own name to the result: that act is very nearly the whole of what craft now means.
What makes this an emergency rather than an irony is that the veto was already dying before the machines arrived to finish it, and everyone who has worked inside an editorial institution in the last twenty years has watched the mechanism of its death. No memo abolishes the veto. The piece is due; the slot is sold; the writer is owed; the editor knows the chapter is not right and says so, and the answer, reasonable and friendly and fatal, is we don’t have time to send it back. The first time, that sentence loses one argument. Repeated for a decade, it rewires the institution: the threshold quietly descends, the editors who keep saying no acquire a reputation for being difficult, the ones who learn to say yes advance, and the institutional memory of what the veto protected thins until a generation arrives that has never seen it exercised and experiences any attempt to revive it as obstruction. Vetoes die on the calendar, scheduled out of existence one deadline at a time, and then forgotten, and the forgetting is what makes rebuilding them so much harder than defending them would have been.
The economics of this erosion are real, which is what makes it tragic rather than stupid. The veto is friction by design. It exists to stop production, and stopping production costs money that the spreadsheet records against benefits the spreadsheet has no column for. The benefit is a compounding asset: the institution’s standing with the people who can tell, which is the only thing, in the long run, the institution sells. Liquidating it pays out steadily, quarter by quarter, with the invoice arriving years later in a currency, trust, that cannot be repurchased at any spot price. A few institutions have understood themselves clearly enough to refuse the trade, and the next chapters visit them. But the aggregate story of two decades of media is the story of that asset being spent, and the machines have arrived at the precise moment the accounts are empty: infinitely fluent production, meeting a culture that has already dismantled most of the seats from which anyone could tell it no.
What happens then is predictable to the point of mechanical. It is a system with the correction mechanism removed, and such systems have a characteristic behavior: the error travels outward from where it started. It compounds. The next chapter traces the four loops by which it compounds (through the platforms, through the audience, through the institutions, and finally through the makers themselves) and why no one in any of the loops ever has to choose the outcome for the outcome to arrive.
Notes (3)
Maxwell Perkins to F. Scott Fitzgerald, November 20, 1924. F. Scott Fitzgerald Papers, Princeton University Library. ↩︎
A fair concession: a machine can do real diagnostic work here. It can flag the blurry character, surface an inconsistency, mark a slack passage, even propose where to cut. That is useful, and the posture this book recommends takes all of it. But flagging a flaw is not the veto. The veto is the authority and the willingness to say no, not yet, and to be answerable when the no was wrong, or when it should have been a no and wasn’t. A model can recommend a thousand times without cost; it cannot stake itself on the judgment, cannot be the one whose standing rides on having stopped the wrong book or passed the right one. Detection is cheap and improving. The willingness to pay for the call is the scarce thing, and it is the only thing the seat is built to hold. ↩︎
Variety; NPR. ↩︎