Nobody Chose This

The strange thing about slop is that nobody has to choose it. There is no constituency for it, no manifesto, no executive anywhere whose stated strategy is worse. Ask anyone in the chain (the platform engineer, the page operator, the reader scrolling at midnight, the writer cleaning up machine drafts for rent money) and every one of them will tell you, sincerely, that they prefer good things to empty ones. Then watch the system they jointly operate fill the world with emptiness anyway. There is a gap between what everyone wants and what the machine of everyone produces, and slop lives in it. It spreads through four feedback loops, each one individually rational, each feeding the others, none requiring a villain. Once seen, they are everywhere: in the feed, in the institutions, and, least comfortably, in the mirror.

The First Loop: Arithmetic

When DiResta and Goldstein mapped the AI-image economy on Facebook in early 2024, the detail that mattered most lay past the shrimp Jesus, in the supply curve. A page operator, many traced to click farms in South and Southeast Asia, could generate a hundred images a day, test them all against the algorithm, and let the platform’s own recommendation engine find the winners, which it obligingly did, pushing the spectacles into the feeds of millions who had never followed the page. The researchers found the platform actively promoting this material, filtering having given way to amplification, because the material performed, and performance is what the ranking system is built to detect. The shrimp Jesus was chosen by an optimization process doing exactly its job, picked by the algorithm with no human at the platform in the loop, fed by suppliers whose costs had fallen to nearly nothing.^[1]

That is the whole loop, and its cruelty is its simplicity. Distribution systems rank what can be measured: clicks, watch-time, shares, the half-second of arrested scrolling. The qualities this book has been tracing (whether anyone stands behind the thing) are not measurable at ranking speed, so the systems are blind to them, a fact of physics rather than malice. And into that blind ranking walks a producer whose marginal cost is zero. A staked human maker can put, at heroic effort, a handful of pieces a week in front of the algorithm’s slot machine. The pipeline can put ten thousand.

This is the birth of the Volume Myth: the belief that ten times more “adequate” content simply increases productivity or choice. It does something worse. It breaks the calibration of the judges. In any system where signal competes with noise, the sheer density of the noise eventually degrades the faculty required to detect the signal, what we might call the atrophy of recognition. We saw this in the collapse of the grading signal during the Chegg crash of 2023-2024: when every student has access to a B+ answer for every question, the B+ stops being a measure of student competence and becomes a measure of platform access. The instructor’s ability to recognize genuine understanding passes from harder to impossible, because the “adequate” machine-output has flooded the zone. The volume liquidates the signal in the very act of seeming to add value. And as it liquidates the B+, it relocates the scarce signal, though not to where the optimist hopes. The reassuring story is that the flood raises the quality bar, pushing the meaningful work up to some tier of excellence the machine cannot reach: the A and beyond, safe above the rising water. Do not lean on that. There is no altitude of quality the machine is guaranteed not to climb to, and a defense that depends on one is a defense with an expiration date stamped on it. What still signals after the flood is the one thing volume cannot manufacture at any price: a person who has staked something on the work and can be found when it fails. The flood does not push the meaningful part of the gradient higher. It strips quality of its power to certify at all, and moves the signal off the artifact and onto whoever is standing behind it.

Take the “Reviewer’s Fatigue” currently hitting the academic and journalistic gates. A peer reviewer for a major scientific journal, who once spent his weekends carefully weighing the nuance of a single methodology, now finds his inbox flooded with twenty papers that all read with the same eerie, A-minus fluency. They clear every specification. They cite the right studies. And yet, after the tenth one, he finds his own faculty of recognition beginning to blur. His attention has slid from the breakthrough he is supposed to be hunting down to a lower and meaner question: is there anyone behind this, or is it competent exhaust. He has been forced to move from an evaluator of insight to a screener for absence, and in the transition, the faculty that could have recognized the next genuine insight is being worn down by the sheer volume of the adequate. The “atrophy of recognition” arrives as a physiological response to a flooded environment, chosen by no one.

At those odds the pipeline does not need to be better; it needs only to be there, at volume, and the volume is the strategy. By April 2026, Deezer was reporting that forty-four percent of all music uploaded to the platform daily (seventy-five thousand tracks a day) was fully machine-generated.^[2] Listeners found their votes being farmed on their behalf: most of those tracks’ streams were themselves fraudulent, bots streaming bot music to harvest royalty fractions, a closed loop of synthetic supply and synthetic demand running inside the machinery built to measure human delight. The scarce, expensive, staked thing competes for slots against an opponent whose ammunition is free. It loses on volume before quality is ever consulted.

The Second Loop: The Diet

Taste is calibrated by diet. The phrase describes the literal mechanics of how perceptual norms work. What feels like an ordinary argument, an ordinary photograph, an ordinary song is set by the distribution of arguments, photographs, and songs you actually encounter, and the feeling updates continuously whether you consent or not. Nobody decides to lower their standards, for the same reason nobody decides to acquire their parents’ accent.

So consider what the first loop does to the diet. The average feed now serves a higher and higher fraction of content from the zero end of the stakes gradient: fluent, complete, committed to nothing. Consumed in volume, it does two things to the instrument doing the consuming. It normalizes the texture of emptiness: the article proposes without asserting, the review that describes without judging, stop registering as failures and start registering as what articles and reviews are like. And it erodes the comparison set: the reader who once would have felt a hollow piece as hollow, because yesterday she read three pieces with someone home in them, now reads it against a backdrop of forty others just like it, and the hollowness reads as genre.

I conceded in chapter one that blind tests keep showing readers unable to distinguish machine prose from human. The inability is itself the output of this loop, a moving quantity rather than a fixed fact about readers, and it is getting worse, measurably, cohort by cohort, as the diet shifts. Discrimination is a trained capacity. The training data, for humans as much as for models, is the environment, and the environment is being repopulated by the very thing discrimination was supposed to catch. The loop reaches past mere failure to detect slop and actively detunes us to it, and the detuning compounds: every reader who stops flagging emptiness sends one more engagement signal that teaches the first loop to serve more of it.

Taste degrading is the visible wound. Beneath it runs a quieter one: even a palate that stays intact does not govern what gets eaten. What the feed optimizes for is not what you would choose in a considered hour but what you click at one in the morning with your judgment worn to nothing, and slop is tuned, at machine speed, for exactly that depleted version of you. So preference and behavior come apart. A person can value the real thing, in the daylight part of himself, and still pour the bulk of his attention into what the machine learned to set in front of the tired animal, and the choice he would make awake never gets cast. It is also the flaw in the hope that scarcity will save us, the forecast that as slop floods in, people will come to prize the human-made thing and seek it out. They may well prize it. Seeking it is a different muscle, and the feed is built to keep that muscle from ever being used. The diet does not only coarsen the palate. It arranges for the palate to stop deciding the meal.

The most concrete casualty is genre. A genre is a contract: an opinion column exists to take a position someone will defend; a review exists to render a verdict someone will own. Strip the commitments at scale and the forms persist like shells: by 2024 it was routine to read a “review” that had clearly never touched the product, an “explainer” that explained nothing because it was compiled to occupy a search result, an op-ed page on which no one could locate an opinion. Readers raised on the shells stop expecting the contract. And a public that has forgotten the contract loses even the ability to miss it, which is the loop’s exit door welding itself shut: demanding the return of something you never saw work is beyond anyone’s reach.

There is a finding from the creativity research that completes the diet argument at a different level, because it operates above any individual’s detuning, in what happens in aggregate when a large group of individuals each make individually better choices. In a 2024 study published in Science Advances, researchers found that while AI-assisted writers were rated more creative and enjoyable by evaluators, their collective output was 10.7% more similar to each other than that of writers working alone.^[3] By anchoring to the same machine-generated “starting ideas,” the writers individually improved their scores while collectively shrinking the pool of human variety. The friction of the blank page was producing something invisible in any single output: the diversity of the whole.

The Third Loop: Borrowed Trust

Chapter six showed this loop’s mechanism in miniature. The veto dies by schedule, not by decree: the piece is due, the slot is sold, we don’t have time to send it back. What turns that local erosion into a systemic loop is what the platforms then do with the eroded institutions’ output. A masthead carries trust accumulated across decades of exercised vetoes; the trust persists long after the practices that earned it have been cost-cut away. So when a hollowed institution starts publishing unstaked volume, that volume travels with the old credibility attached, ranked higher, shared wider, believed more than anonymous slop could ever be. The reader who has rationally learned to discount the random content mill has no defenses against the brand her family trusted for fifty years; the trust is what is being spent. Sports Illustrated’s machine-written product reviews, which we will dissect properly in the next chapter, did their damage through the franchise’s history rather than despite it: seventy years of accumulated belief, liquidated through a vendor pipeline at affiliate-revenue rates. Each such liquidation feeds the other loops twice over: it pours credentialed slop into the distribution pool, and it teaches another tranche of readers that even the trusted names are hollow now.

The Fourth Loop: The Mirror

The philosopher Albert Borgmann, writing in 1984 about technologies far gentler than ours, drew a distinction worth handling. Some things in a household, he observed, are focal: they gather skill, effort, and attention around them: the hearth that must be tended, the instrument that must be practiced. Others are devices: they deliver the same commodity (warmth, music, dinner) while hiding the whole apparatus of its making behind a button.^[4] Most device-substitution is pure gain. Central heating beats hauling logs; no one’s character is improved by being cold while he splits wood, and the thermostat hands back an afternoon for something that matters more. Autocorrect took spelling, and spelling was a perfectly good thing to lose. We each stand on a mountain of inherited competence we will never personally hold, and refusing to stand on it, reinventing every wheel by hand, amounts to waste dressed up as virtue.

Borgmann’s real point is narrower and sharper. A device sometimes removes not only the labor but the competence to judge the commodity it hands you, and that single loss is the only one worth losing sleep over. Matthew Crawford, a quarter-century later, found it under a motorcycle. The man who has never repaired anything lacks more than repair skills; he lacks the judgment to know whether the shop actually fixed his bike, and has to take “it’s fixed” on faith.^[5]

So the question to ask of any outsourced task is not am I losing a skill. You are, all the time, and almost always it is fine. The question is am I losing the ability to tell whether the result is any good. Hand arithmetic to the calculator and the answer is no: you keep the judgment that the restaurant bill looks wrong, and the calculator never lies to you. Hand over writing, or any knowledge work whose product is an argument, an analysis, a verdict, and the answer is yes. Writing was the focal practice of knowledge work, and the models have turned it into a device. The button delivers the commodity: a draft, a chapter. But the apparatus hidden behind it was the thing that built the judgment, something far from the drudgery of hauling logs: the confrontation with the gap between what you meant and what the sentence says, the discovery that the argument you believed has a hole in paragraph four, the slow accretion across ten thousand such collisions of the internal standard you measure work against. Outsource the heating and every faculty that mattered survives. Outsource the writing and the faculty you lose is the one that could have told you whether the writing is good, the one a tool cannot hand back, because checking its work requires the very judgment you traded away for the convenience. The standard was the labor itself, gone with it rather than tucked away in storage.

A few months of approving fluent output, and it begins, like any unused discrimination, to flatten. The approver still reads the draft. It still seems fine. That is exactly what the atrophy feels like from the inside: everything seems fine, at lower and lower altitudes of fine. This is the ultimate cost of stakelessness: the erosion of the human capacity to be answerable for quality, because the standard itself has been outsourced. The dignity of the maker is replaced by the passivity of the approver.

And here the fourth loop closes into the first three with a neat, terrible click. The maker whose feel for the difference has gone dead cannot prompt for the difference, cannot edit toward it, cannot veto its absence, cannot even miss it. Her output drifts toward the floor of what the tools produce unsupervised, which raises the slop fraction in the distribution pool (loop one), which further detunes the audience (loop two), which further weakens any institutional case for expensive judgment (loop three), which pushes more makers from making to approving (loop four), around and around, no villain anywhere, everyone optimizing, everything getting worse.

---

Four loops, one engine, and the engine has a single fuel: every loop is powered by the removal of a cost. The platform loop runs because production became free. The audience loop runs because consumption requires nothing. The institutional loop runs because the veto was expensive and the spreadsheet found the expense. The maker’s loop runs because drafting was effortful and the button took the effort away. The system is working perfectly, doing what every actor asked of it. It removes friction, eliminates cost, and makes everything easier, and the cumulative output of all that ease is the engulfing tide of nothing this book opened with, because the costs being removed were the very thing holding it up. The friction was the filter. The effort was the standard. The expensive judgment was the quality. Remove them and the structure goes hollow rather than light, and hollow structures stand, looking exactly like sound ones, right up until weight arrives.

Each loop makes a claim that could fail. The first is not operating if cheap synthetic supply stops gaining distribution share, if the platforms rank provenance over performance and the volume play stops paying. The second is not operating if discrimination holds steady or improves as the diet shifts, cohort by cohort, instead of dulling. The third is not operating if trust in hollowed brands tracks their actual practices rather than their accumulated names. The fourth is not operating if approvers who lean on the tools keep their edge against a held-out standard instead of drifting toward the floor. None of these is guaranteed to come out the way the chapter predicts. That is what makes them loops and not articles of faith.

“Nobody chose this” needs one qualification before it can be trusted, because in its loose form it converts decisions into weather. No one chose the loops; people chose, repeatedly, the things the loops run on. A private-equity firm decides to strip-mine a fifty-year-old masthead for its domain authority and run it on machine copy. A platform decides which ranking signals to optimize and which to ignore, and the choice to reward engagement over provenance is a choice, made in a meeting, by people with names. A lab decides to ship a model at a given capability and a given guardrail, on a given date, against a given internal objection. The structural account and the agency account are both true at once: the incentives are real and impersonal, and specific actors keep deciding to lean into them rather than against. A book that builds a vocabulary of answerability and then exempts the most answerable parties has misplaced its own argument. The loops are the river; the firms above chose to dam it for power.

One implication of the systemic account this chapter has been making (the loops run on structural incentives, each one individually rational) is that it seems to leave no room for the individual. If the flood is nobody’s fault in particular, then keeping the standard is nobody’s work in particular, and the only lever is policy, or platform design, or some other thing that acts at the level where the loops live. That reading is wrong. The loops are structural, yet they stop short of total. The system has no mechanism to prevent an individual from opting out of any one of them, at any time, and the person who does carries something the loops have no way of producing: a standard maintained when no one is enforcing it. The maker who keeps a bar above adequate when the machine delivers adequate for free; the editor who sends a piece back when the schedule says ship it; the reader who pays for the staked version when the unstaked one is right there and indistinguishable: each of these acts is economically irrational in the short run, and none of them is heroic. They are simply the thing someone still decided to do. Every institution of quality that still works is downstream of someone who made that decision repeatedly, over a long time, before the institution existed to reward it. The keeper of the standard, far from being a relic of a better era, is most of the time the only reason there is an era that followed.

Take two people with identical tool access: same models, same subscriptions, same prompt techniques. The first uses them to generate landing page copy that reads like expertise, to spin aggregated sources into what looks like original research, to populate a product site with testimonials that feel like they came from customers. Every output is fluent. None of it is false in any checkable particular. The whole apparatus is pointed at a single end: extracting money before anyone notices that no one is accountable for the claims. The AI found this person’s instinct already in place and merely removed the friction between the instinct and its expression. Before the models, the gap between wanting to grift and being able to sustain one was wide enough that a lot of grifts never got off the ground. The tools closed it.

The second person (call the type the infrastructure builder, the one with a ten-year project and a public record of having been right and wrong in approximately the proportions you’d expect) uses the same models for something else entirely. The AI drafts, and she edits. It generates candidate arguments, and she kills the weak ones and expands the sound ones. It handles the work that was previously friction between her and the thing she was trying to build, and what she is building has stakes in the most literal sense: a reputation that accumulates against it, readers who can compare her predictions to what actually happened, work that will either be useful or won’t and can be checked by anyone who cares to look. The output from this person is also fluent. It is also fast. The difference is not detectable in any single artifact. You have to look at what’s behind the artifact: the real thing being built toward a real end, or the performance of a thing with nothing behind it.

This is what “AI scales your direction of travel” means, made personal. The moral accountancy is more banal than the tool making good people better and bad people worse. The tool removes friction, and friction was doing a job. For the builder, the friction being removed was between intention and execution: the draft she couldn’t get down fast enough, the research she didn’t have time to synthesize. Removing it gets her to the destination sooner. For the grifter, the friction being removed was between intention and consequence: the gap that limited how fast the extraction could run. Remove it and the extraction scales. Both people are using the tool exactly as designed. What the tool cannot change is the direction: what they were already running toward, and what happens when they arrive faster.

This matters for the chapter’s argument because the loops are structural but they operate on individual decisions. The platform loop, the diet loop, the institutional loop, the mirror loop: all four have the same shape as the grift: systems that remove friction from a direction of travel already in motion, extracting value in the gap before consequences catch up. What interrupts the loops is the same thing that separates the two users above: someone who decides to be accountable for the destination. The work falls to a person rather than to better detection. Someone for whom arriving faster matters because the destination is real.

Which means the loops have a single point of vulnerability, and it is not the one the discourse keeps reaching for. They cannot be stopped by detection tools, which lose the arms race a model generation at a time; or by labeling laws, which tax honesty and exempt the indifferent; or by waiting for audiences to revolt, since loop two is busy dismantling the faculty a revolt would require. The loops are powered by costlessness. The only thing that interrupts them is the reintroduction of cost, somewhere, by someone, on purpose: an institution that pays for the veto when the spreadsheet says not to; a maker who keeps drafting when the button is right there; a reader who pays, in money and attention, for the staked thing when the free thing is adjacent and adequate. Every one of those acts is economically irrational in the short run. The institutions in the next chapter have been committing them, deliberately, some for sixty years, and the strange news from their balance sheets is that the irrationality is starting to pay.

Notes (5)

1. Renée DiResta (formerly Stanford Internet Observatory) and Josh A. Goldstein (Georgetown CSET), “How Spammers, Scammers and Creators Leverage AI-Generated Images on Facebook for Audience Growth,” CSET, March 2024. Roughly 120 AI-image mills mapped. ↩︎

2. Deezer Newsroom; TechCrunch, April 2026. ↩︎

3. Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 10, eadn5290 (July 12, 2024). DOI: 10.1126/sciadv.adn5290. The study used 293 writers and 600 evaluators; writers individually improved their creativity scores using AI, but the collective similarity of their stories increased by 10.7% compared to the control group. ↩︎

4. Albert Borgmann, Technology and the Character of Contemporary Life (1984). ↩︎

5. Matthew B. Crawford, Shop Class as Soulcraft (2009). ↩︎