SLOP

Chapter Four

The Floor and the Attention

It is the spectator, and not life, that art really mirrors.

— Oscar Wilde, The Picture of Dorian Gray (1890)

Start with the easy tests. Ask a model whether the Battle of Stalingrad ended in a German surrender, and it will tell you: yes, January 1943. Ask who won the 1948 World Series, and the Cleveland Indians beat the Boston Braves in six. Ask about compound interest or the Krebs cycle, and the answer is organized, correct, and clearer than what a human professional would produce on an average day, the key word being average. A professor has average days; the model does not. The cell biology teacher who covered that unit for fifteen years still delivered it differently after a bad night, a difficult conversation, a Tuesday in November when she was somewhere else in her head.

For decades, we assumed this would be the hard part. Getting the facts right was supposed to be the summit. Fluency was the mystery for our grandchildren. Both assumptions are now rubble. The machines took the two qualities our whole evaluative tradition was built to measure (is it true? is it well made?) and turned them into commodities almost overnight.

Supporters cheat by pretending the conquests are total; defenders of human craft cheat by pretending they are fake. The evidence buries both. The conquests are real, and the conquered territory turns out to have been a place quality had already left.

Take the conquest of truth first. Its failures concentrate exactly where checking is hardest. Ask about Stalingrad and the model is grounded by ten thousand books. Ask about a minor Byzantine official or a subclause of maritime law in a lesser known North African port, and the confident register continues, but the signal of how solid the ground is drops away. Calibration has improved on aggregate; recent models hedge and abstain more often than their predecessors did, and the selective-confidence research is real. The improvement runs along the average, though, and the specific query that crosses from the well-trodden into the speculative is exactly where the score is least to be trusted. Often the model still maintains the same confident register whether it is citing Stalingrad or confabulating a Byzantine footnote, and both the system and the user are left guessing which mode they are in. The failures land where checking costs more than asking.

Concede the optimistic case in full, because the argument does not need to deny a word of it. The confident-wrong answer that defined the early models grows rarer with each release, and there is no reason to think the trend stops. A model that flags its own uncertainty cleanly and confabulates a fraction as often as its grandparent did three years ago is a near-certainty, not a fantasy, and a reader who bet against it would lose.

But that improvement is the opposite of a rescue. When the machine was wrong a third of the time and brazen about it, the crudeness worked as a warning label; you knew to keep your guard up. Drive the error rate down and polish whatever is left, and you remove the last thing keeping anyone alert. The mistakes that survive are the ones that survived because they are hardest to catch: the plausible Byzantine footnote, the subclause that reads exactly like real maritime law, the confident figure in the one field where you had no way to check. A calibrated confidence score helps right up until the case that matters is the case where the score was wrong, and a score is not a someone. It cannot be deposed, cannot be sued, cannot be embarrassed when the eighty-percent answer turns out to be the twenty. The better the machine gets at truth, the more completely the only question left is the one this book keeps returning to: behind this answer, in this case, is there anyone who pays if it is wrong. Clearing the truth bar does not retire that question. It strips away everything that used to stand in for it.

A second failure runs alongside the first: true-but-generic. The patient asking about chest pain needs a doctor to notice she mentioned it as an aside, after the trouble sleeping, and that her family history makes the “statistically musculoskeletal” answer a dangerous permission to stay home. The doctor’s scarce skill always lay past knowing the literature, in knowing which knowledge this moment requires and being answerable for the choice. The models have flooded the literature and run dry at the choice.

Wikipedia works, and it is worth conceding why. It works because its domain is the cheap-to-verify. It is an excellence built on process and distributed accountability: editors accumulate standing within its governance, disputed claims carry the visible marks of arguments fought to resolution, and being caught inserting nonsense carries real costs. Wikipedia has stakes; they are communal and process-verified rather than attached to any individual byline, which is why it works for the cheap-to-verify and why it cannot do what a named expert does when a claim must have a single address. This book’s thesis is fine with that. Wikipedia thrives outside our boundaries because its domain does not require individual testimony. The mushroom guide does.

The machines are also producing wonders, like DeepMind’s AlphaFold, which predicted structures for two hundred million proteins.[1] It is trusted because it was checked against decades of lab data, ships with confidence scores, and was released by a named institution that staked its reputation on the tool. The hopeful future and the slop crisis are the same mechanism: value flows wherever a claim can be tied back to something that answers for it.

The boundary generalizes: most of what gets produced can stay well short of a Monet. A grocery list wants accuracy and nothing more. Slop is the failure to supply stakes when the leaning happens (the stakes in the medical answer, the presence in the witness), a failure of placement rather than of lavishness, while the machine fluently supplies the appearance of both with little discrimination about when it matters.

Accuracy has become a floor. When every essay is spelled correctly, spelling tells you nothing. When every answer at least has the appearance of accuracy, then accuracy also tells you nothing. This is how floors work: they neutralize the old signal while exposing what was built on it. For five centuries, a document that was accurate and well-made was, reliably, a document somebody had committed to, because accuracy and craft were hard enough that their presence proved a person. The floor arrived and that proof collapsed. A floor of near-universal apparent accuracy means the question is this right stops sorting anything. The question that starts sorting is the older and harder one: who will answer if it is wrong? The machine raised the floor. The stakes question is what is left standing above it.

Then the second conquest: beauty. In April 2023, millions enjoyed “Heart on My Sleeve,” a song featuring synthetic versions of Drake and the Weeknd.[2] The discomfort was that it was good, and the source was nothing.

We should stop lying about machine beauty. If the sentence catches your breath, the catch happened. You can be told later that no one wrote it, and that changes what it stands for, but it doesn’t un-catch your breath. The machines have manufactured the appearance of beauty, and the appearance of beauty is the experience of beauty.

So what was beauty doing for us before it was rentable? It was evidence of attention.

The opening line of Camus’s L’Étranger is four words: Aujourd’hui, maman est morte. The first English translator, Stuart Gilbert, rendered it “Mother died today”: accurate by any dictionary while also insufficient. Maman is not “Mother”; it is the child’s word, warm and unguarded, surviving in the grown man’s mouth on the day of her death, set against the flat finality of est morte. The whole novel’s emotional problem sits compressed in a register clash that “Mother died today” flattens. It took until 1988 for Matthew Ward to make the informed call: leave it. Maman died today.[3] Keep the foreign word, keep the warmth and the strangeness, let English readers feel the thing French readers feel. Hundreds of pages of competent translation, and the entire art is visible in that one refusal to settle: a translator holding the original’s performance in one hand, the target language’s resources in the other, and declining every adequate option until he found the one that did the work, at the cost of a choice he would have to defend.

Design tells the same story in cleaner form. In 1972 the designer Massimo Vignelli gave New York a subway map that riders spent seven years hating (the geography distorted, Central Park squashed, the water the wrong shape) and that designers have spent fifty years studying, because every one of its distortions was a decision in service of the only question a subway map exists to answer: which train, which transfer. Vignelli attended his creation to what the thing was for, and sacrificed everything else to it, including being liked. The city retired the map in 1979 and replaced it with a friendlier one that is harder to use.[4] Both maps are competent artifacts. Only one of them is the record of someone seeing the problem whole and paying a public price for the clarity, and the difference between those two kinds of competence is what the word beauty, in the craft sense, was always tracking.

This is what the beauty axis was measuring, all along: evidence of attention, not polish. The philosopher Iris Murdoch, borrowing the word from Simone Weil, argued that attention is the fundamental moral and aesthetic act: the discipline of seeing a thing as it actually is, against the constant pull of seeing what is convenient, flattering, or expected. Her famous example involves no art at all: a mother-in-law, M, who finds her son’s wife vulgar and tiresome, and who, suspecting her own jealousy, goes back and looks again, deliberately, until she sees a young woman who is not vulgar but unpolished, not tiresome but young.[5] Nothing in the world has changed; the work was all interior, effortful, against the grain of the self. Murdoch’s claim is that this work, looking until you see truly, is the root of all making worth the name. The painter attends to the apple until the apple is internalized, and becomes the apple his intuition knows rather than the apple as apples are supposed to look, then reaches for the canvas. Ward attended to maman until the word’s actual weight, not its dictionary entry, reached the English. When a made thing strikes us as beautiful, what we are responding to, on this account, is the visible record of someone having looked that hard. Beauty is what attention leaves behind.

Murdoch’s own argument creates a tension. The attention she describes is a practice of unselfing: M succeeds by removing her own preferences from the encounter, getting out of the way of what is actually there, rather than by committing herself to a verdict. That sounds like the inverse of what the book is arguing, a matter of getting yourself out of it rather than putting your name on the line.

But the stake and the unselfing are the same act seen from different sides. Having your name behind the work removes the option of settling for the convenient version; the accountability you cannot escape is what makes it possible to come back and look again, without the exit of a comfortable dismissal. M could only see the daughter-in-law once dismissal stopped being the safer option; the commitment that made the second looking possible was accountability itself. The machine’s limit is a matter of exposure rather than processing power. It cannot be wrong in a way that costs it anything, and the possibility of that cost is what makes attention more than simulation.

Take the Staked Designer in a modern branding firm. She is presented with twenty logos that all clear the specification of the brief: they are blue, they use the right font, they are “adequate,” and whether they were made by a human or not, her conclusion is the same and refuses them all. She goes back to the drawing board for a week because she noticed something about the way the client’s product actually feels in a user’s hand, something the brief didn’t say and the model didn’t see. Her refusal is economically irrational; the client would have been “pleased” with the A-minus version. But she pays the cost of the extra week because she is attending to the reality of the work. When the final logo arrives, its “beauty” lives past its symmetry, in the fact that it is the record of that one person having seen a truth that others had missed.

The relativist has been waiting through all of this, with the objection raised whenever beauty is discussed as if it were real. But taste is subjective. One viewer weeps during the film, another walks out; one reader’s profound poem is in another’s pile for the second-hand store. If beauty is only preference, then the machine that manufactures pleasing has taken nothing from anyone, because there was nothing there to take, only opinion, and opinion is always cheap and limitless. Everything here depends on beauty having been a real signal, and a real signal cannot be pure preference.

The answer is old, and the two philosophers who gave the first attempt at an answer were nobody’s sentimentalists. David Hume, in 1757, looked straight at the subjectivity of taste and declined to surrender to it. Some verdicts, he noticed, do not actually vary: across centuries and languages the same handful of works keep being judged great by the people best equipped to judge them, and the ranking is stable enough that anyone who called a hack the equal of a master would be exposing a defective taste rather than expressing a different one. The standard of taste lives in the converging verdict, over time, of qualified judges, somewhere past both the object itself and any tally of opinions: people with practiced perception, wide comparison, freedom from prejudice, the delicacy to catch what the careless miss.[6] Kant added the stranger half. When you call something beautiful, you reach past reporting what you happen to like, the way you report liking cilantro; you are claiming that others ought to agree, making a demand on them you cannot prove and cannot stop making. A judgment of taste reaches for a universal assent it can never fully secure, and that reaching is what separates this is beautiful from this pleases me.[7]

What both men were describing, in the idiom of their centuries, is that taste is trained, communal, and answerable, which is to say, it has the structure this book keeps finding under the word stakes. The disagreement between two viewers shows the ordinary working of a standard that lives in trained judgment rather than in the artifact, one that converges, slowly and never completely, among the people who have actually done the looking. The machine can manufacture the surface that trips the response. What it cannot do is join the community of judgment that the response answers to: it has not looked, it cannot be answerable for the looking, cannot be a qualified judge whose agreement with other judges is the thing that makes a verdict more than a vote. The beauty it makes is real in the eye and empty in the chain: the same split, arriving on the beauty axis, that the truth axis already showed.[8]

An objection arrives here that the technology is already running in practice: AI systems are being placed on art-contest juries, curating acquisitions, ranking submissions.[9] If the machine can serve as judge, does it not join the community of judgment?

The answer turns on what judging requires. A judge who cannot be shown to have been wrong, persuaded by a colleague, or held accountable to a body whose esteem matters to her is merely running a preference algorithm. The community of aesthetic judgment converges through the mutual accountability of those preferences, something preference-aggregation alone can never reach: judges who can be wrong in ways that cost them something in the room. An AI judge can prefer. It cannot answer for the preference, and a verdict that cannot be answered for is not yet a verdict in the relevant, practical sense.

Come down from the abstract to a single case. The claim that beauty is the residue of a real person attending to a real loss is easiest to see where we know the life behind the work. Elizabeth Bishop wrote “One Art” as a villanelle, a tight form built on two lines that keep circling back, and used the repetition itself as a discipline for bearing loss. The refrain insists that the art of losing isn’t hard to master, and the poem tests that claim against a rising scale of things lost: keys, houses, cities, a continent, the controlled voice holding through stanza after stanza, until the last one, where it faces the loss of a person and the composure finally cracks open in a parenthesis: though it may look like (Write it!) like disaster.[10] That self-command (the poet ordering her own hand to set down the word her whole strategy of composure was built to avoid) is the most famous two words in the poem, and their force is inseparable from what they are evidence of: thirty years of practiced restraint meeting the one thing it could not contain. The beauty is the biography breaking through the form.

A language model can produce that shape. Trained on every poem that ruptures, it can rupture on schedule, and a reader who does not know the provenance may well catch her breath. The output can be beautiful; we have conceded that and the concession stands. What the output cannot be is evidence. There was no composure, no thirty years, no thing being contained; only a system that has learned that poems sometimes do that. The surface is present; the attention the surface used to certify is absent. The signature is perfect and there is no signatory.

For the reader, in the moment, the two are indistinguishable. But notice what this does to beauty’s old job. For all of human history, a beautiful surface was a reliable certificate, because the surface was expensive: you could not fake the years it took to write like that, see like that, choose like that. Beauty meant somebody looked hard at this. The models have done to that certificate exactly what they did to accuracy’s certificate: the quality survives, severed from what it certified. There are now two kinds of beautiful objects in the world, the record of a perception and the performance of one, and no test on the object tells you which you are holding.

This inability to distinguish them on the surface has a behavioral consequence the culture is already working through. Whenever a style becomes the machine’s style, it stops functioning as the human’s tell. Writers who had their own ways of marking thought (the em-dash as a rhythm-break, the sentence fragment deployed with calculation, the deliberately unusual vocabulary choices, the specific cadence of a long sentence shattering into a short one) find those signatures appearing in every generated paragraph and quietly drop them. The mark that certified presence becomes the mark that certifies nothing; its absorption into training data is its funeral. This has always happened to expressive forms (slang gets adopted and loses its force, genre conventions calcify from fresh to clichéd), but the speed is different now. Features that took decades to travel from distinctive to default now make that journey in months.

The response comes in two directions. Some creators reach for new tells: stranger rhythms, more idiosyncratic choices, whatever the model cannot yet do, only to watch the next training cycle close the gap. Others go the opposite direction and stop trying to signal humanness altogether, treating the machine’s apparent impossibilities as a flex: output so metrically precise or so ornately structured that it reads as a different kind of proof: no longer a human wrote this but someone with serious ambition marshaled serious resources, and look what they made. Both responses are answers to the same pressure. Neither restores the old certificate, because the certificate always lived past the style itself, in the cost of having developed it, the years of reading and making that gave the mark its trace of a real mind. The machine can acquire the trace. It cannot acquire the formation that produced it. What survives the cycle is the pressure itself, which keeps generating new tells as each old one is absorbed.

Push the concession to its limit, because the decade is going to push it there anyway. In March 2016, in a Seoul hotel ballroom, a machine called AlphaGo played the thirty-seventh move of its second game against Lee Sedol, then one of the world’s strongest Go players. The move looked like a mistake. Commentators apologized for it on the broadcast; the system itself estimated that a professional human would have chosen it about once in ten thousand games. It won the game, and it has since entered the small canon of moves that players study the way poets study a perfect stanza. Fan Hui, the European champion who had already lost to the machine in private, groped for language and landed here: “It’s not a human move. I’ve never seen a human play this move. So beautiful.”[11]

He spoke plainly, and with no one there to flatter, flattery was beside the point. The beauty of move 37 was real, public, and durable, and no human mind conceived it. Hold that fact and run the thought experiment the technology is now preparing to run for real. Somewhere a system operates with no prompt worth the name and no person in the chain: it generates a film, end to end, unwatched by its own pipeline. Suppose the film is good. Suppose it is wonderful: people send it to each other saying you have to see this, it gathers a hundred million viewers, and there is no director to interview, no one to call up the aisle on awards night. By the letter of this book’s definition that film is slop: output from the far end of the gradient, with no one staked on it at any point. And the letter, applied that way, would be wrong.

Ask what the film actually promised you. A piece of entertainment makes one wager: that it will repay the hours it asks for.[12] That wager is tested in the watching: the collision and the consumption are the same event, which places the whole genre on the cheap-to-verify side of the boundary this chapter drew at Wikipedia. The mushroom guide’s betrayal arrives weeks after the reading, over dinner, in a body. A bad film’s betrayal arrives in the second act, instantly, at no cost beyond the evening you were spending anyway. Where checking is free, stakes were never the strand the situation was leaning on; the film’s one contract was the experience, and the experience audits itself in the watching.

So the honest category for the authorless masterpiece is nature. Humanity has always been moved by what no author made; we have wept at sunsets, at canyons, at the geometry of a shell, all of which stand alone, and all of which we call beautiful rather than slop. What the generators have done, at the far autonomous end, is return the made thing to the condition of the found thing: beautiful the way weather is beautiful, and mute the way weather is mute. You can love it. What it withholds is an answer to what it meant, someone to thank, a promise about the next one, because there is no one whose standard it expresses and no career it puts at risk. Move 37 committed nobody to anything, and the film by no one carries no promise of a next film. Found beauty is a gift without a giver, and the gratitude it earns is real and has nowhere to land.[13]

The conquest itself, though, is never authorless. The film spreads by recommendation: one person telling another trust me, this is worth your evening, each share a small wager of credibility placed before people who remember bad tips. The film carries no stakes, and its distribution is made of almost nothing else. The audience, not the maker, supplies the testimony, the way Dublin’s crowd supplied everything real on O’Connell Street. That is the era’s strange new commons: oceans of unstaked wonder, navigated by chains of small human vouching.

The boundary, then, again, is: entertainment can be authorless; testimony cannot. Recut that same film as a documentary, let it claim anywhere that this happened, and the void behind it starts to matter again, instantly, with bodies attached. The same artifact crosses from the cheap-to-verify to the expensive-to-verify, and the missing answerable someone stops being a curiosity and becomes the whole question.

So both of the measurable axes end in the same place. Truth conquered, and revealed as a floor. Beauty conquered, and revealed as a certificate now forgeable at scale. Everything our evaluative tradition could point to in the artifact itself (the claims check out, the craft is superb) the machines can now supply, which means the artifact itself can no longer answer the question we were always asking of it. The question was: can I trust this? did someone stand behind this? if I lean on it, who falls? For hundreds of years the artifact carried the answer bundled inside its scarcity; that bundle is now broken.

The argument has not been that the machines produce imitation truth and fake beauty while humans produce the real thing. It has been more specific: the appearances of truth and beauty are separable from what makes them matter in the world, and stakes are what the other two become when they are fully arrived at rather than merely approximated, not a third strand supplementing them. Testified truth is what a claim is when it has a real address and someone to answer for it, which is more than accurate-plus-accountable. Committed beauty is what craft is when it is the record of an attention rather than a performance of one, which is more than polished-plus-intentional. The machines have demonstrated that the surface of each can exist without the structure that made the surface meaningful. What they cannot demonstrate, because they cannot do it, is that the surface was ever what the world was actually asking for.

What’s left is the third thing, the one that was never in the artifact at all. It lived in the relation between the work and a person who could be ruined by it, and it is the only one of the three the machines have not touched, because it is the only one that cannot be generated, for the same reason a forged signature cannot vouch for itself.

What that third thing is made of, precisely and structurally rather than as a slogan, is the next chapter, and it begins with the greatest forger of the twentieth century selling a fake masterpiece.

Notes (13)

  1. Google DeepMind’s AlphaFold2 won the CASP14 protein-structure competition in 2020; in 2022 DeepMind released predicted structures for roughly 200 million proteins via the AlphaFold Protein Structure Database. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry (with David Baker, awarded the other half for computational protein design) for the work. DeepMind; The Nobel Prize. ↩︎

  2. Posted April 2023; drew millions of streams before being pulled by Universal Music Group. Variety; CNN, 2023. ↩︎

  3. Camus’s L’Étranger (1942); Stuart Gilbert translation (1946); Matthew Ward translation (1988). ↩︎

  4. MoMA; “1972 Vignelli Map of the New York City Subway System.” ↩︎

  5. Iris Murdoch, The Sovereignty of Good (1970). ↩︎

  6. David Hume, “Of the Standard of Taste” (1757). ↩︎

  7. Immanuel Kant, Critique of the Power of Judgment (1790). ↩︎

  8. The hard version of the objection deserves its full strength. A thoroughgoing subjectivist (the tradition runs from de gustibus non est disputandum through the logical positivists, for whom “this is beautiful” is just “hooray for this,” an emotive noise with no truth value) denies there is anything to converge on at all. If that held, the appeal here to a community of judgment would be empty: only preferences, equally valid, all the way down. Two things answer it. First, almost no one lives as though it were true. The same people who say beauty is merely subjective do not actually rate a supermarket greeting card the equal of Rembrandt, and the asymmetry in how seriously they hold the two verdicts is the standard reasserting itself against the theory. Second, the convergence is observable: across centuries and languages, qualified judges keep ranking the same works highly, and the stability of that ranking is a fact about the world that pure preference cannot explain. None of this requires beauty to be a property of the object the way mass is. It requires only what Hume claimed: that trained perception, given time and comparison, converges more than chance would predict, which is all that “a real signal” ever meant. ↩︎

  9. The practice is real, and already instructive. Beauty.AI, in 2016, billed itself as the first international beauty contest judged entirely by machines: roughly six hundred thousand people submitted selfies to a “robot jury” of five algorithms scoring symmetry, wrinkles, and the like. The result was a parable about this chapter’s point: nearly all the winners came out light-skinned, and the project’s own chief science officer conceded the training data had been skewed. Curation has followed the same path. In September 2023 the Nasher Museum of Art at Duke mounted an exhibition whose themes and checklist were generated by ChatGPT, and the 2023 Helsinki Art Biennial folded machine curation into its program. These systems can rank and select. What none of them can do is be answerable for the ranking. TechSpot and Vice (Beauty.AI, 2016); MuseumNext (Nasher Museum, 2023). ↩︎

  10. Elizabeth Bishop, “One Art,” collected in Geography III (1976). ↩︎

  11. The move is commonly referenced as “Move 37.” “AlphaGo versus Lee Sedol.” ↩︎

  12. “Repay” is deliberately broad. The currency varies (laughter, suspense, catharsis, the good cry, awe, the satisfaction of a plot that locks shut) and a single film often pays in several at once. What unites them is the timing: each is delivered and judged in the watching, which is what keeps the whole category on the cheap-to-verify side of the line. A work that asked to be trusted for something it could deliver only later, a documentary’s facts, a manual’s instructions, would be making a different wager and would need a different strand to back it. ↩︎

  13. A fair objection: a canyon makes no claim, but a model can state a truth about the world, and a true statement is not mute the way weather is. The asymmetry holds anyway. Aesthetic output asks nothing of its source; we can be moved by beauty no one meant, the way we are moved by a coastline. A truth-claim is different in kind, because a claim has the shape of something a claimant asserts, and assertion is exactly what the machine does not perform: it emits a sentence that corresponds to the world, or fails to, with no one standing behind the correspondence. So a model’s true sentence is closer to found beauty than it first looks, an accurate artifact thrown off by the compressed reasoning of millions of human authors, useful and even valuable, but not an act of telling the truth, because no one is answerable for its being told (see the introduction’s note on accuracy versus truthfulness). We can take the accuracy as a gift, the way we take a vein of ore. What we cannot do is thank it, or hold it to its next claim, or ask it how it knows. Truth without a truthteller is found, not testified, and that difference is most of this book. ↩︎