Chapter One
A Promise and a Body
The pictures showed a promise. Cascades of chocolate under glass domes. A garden of candy-colored growth dissolving into golden light, lollipops the size of streetlamps, a purple-coated figure presiding over places the website called an Enchanted Garden, an Imagination Lab, a Twilight Tunnel. The pictures had the particular lushness that image models produce when nobody reins them in: every surface gleaming, every color saturated past natural norms. Some of the text on the website had that telltale shimmer of words arranged by something that does not read: a pasadise of sweet teats, one line offered. Catgacating. Exarserdray lollipops. Tickets to Willy’s Chocolate Experience, Glasgow, February 2024, cost: thirty-five pounds. Thousands of parents looked at those pictures and did what parents do: imagined a child’s face in that golden light, and paid.
What the tickets bought was a sparsely carpeted warehouse on an industrial estate in Whiteinch, a handful of plastic props, a small bouncy castle, and actors who had been handed a script of AI-generated dialogue while trying gamely to perform; including a character no one had heard of: the Unknown, a silver-masked villain who lived in the walls and, in the event’s single accidental stroke of genius, terrified the children present more efficiently than any designed experience that year. A girl in a green wig handed out a single jelly bean per child, the supply being finite in a way the pictures had not been. By early afternoon, the families in the queue called the police. The event was shut down by evening. Within a week the Unknown was a global meme, the organizers were refunding tickets under threat, and “Glasgow Wonka” entered the language as shorthand for a new and specific kind of betrayal.[1]
What failed in that warehouse is less obvious than the jokes made it look. The images were not bad. The images were, in the only sense images themselves can be judged, accomplished. They did exactly what promotional images exist to do: make people want the thing. What failed was the one thing the images were silently presumed to carry: a connection to anyone who had committed to making them true. A photograph of an event is, among other things, an implicit promise: someone stood here, this exists, you can stand here too. No camera had been anywhere near those pictures, and nothing stood anywhere. They were generated by a system with no knowledge of the warehouse and deployed by people whose plans had never caught up to what the images had promised; every pixel of their beauty was beside the point.
Eight months later, on the last day of October 2024, thousands of people lined O’Connell Street in Dublin for a Halloween parade. A website called myspirithalloween.com (registered abroad, stuffed with AI-assembled event listings, monetized with ads) had announced a 7 p.m. parade and attributed it to Macnas, a real and celebrated Irish street-theater company, with persuasive photographs from real, previous parades. The crowd assembled. The crowd waited. There were costumes, children on shoulders, the full apparatus of civic anticipation. It took the national police confirming there was no parade before the street emptied out. The man behind the site, reached by reporters afterward, said his team was “highly embarrassed” and that they made a mistake. Nobody set out to gather ten thousand Dubliners on a curb. Nobody had set out to do anything. That was the point. A system for generating plausible event listings had generated a plausible event listing, and the listing had done what plausible things do: it was believed by bodies who showed up.[2]
From the outside, both were comedy. Glasgow became a meme because it was meme-shaped: the Unknown found a fan base, a pasadise of sweet teats found T-shirts, and most of the planet met the whole affair as delightful farce. They were right to. Almost no one was badly hurt. A spoiled Saturday, thirty-five pounds refunded, a child frightened and then carried home. The stakes were low. What should stop the laughter is the next thought: that the machine behind the farce, indifferent in the same way to whether anyone stood behind its promises, is the same machine now turned to work no one laughs at, the kind people act on before they find out. Glasgow is slop in its comic register, where the bill comes to nothing. The rest of this book follows the same mechanism into rooms where the bill comes due.
Those crowds, Glasgow’s and Dublin’s, will keep reappearing in this book, because people are going to stand on that curb again, in a hundred forms. The feed promises, the body shows up, but nothing is there. The gap between a generated promise and a staked one has a texture, and the texture was being felt long before it had famous examples, in a thousand smaller doses that never made the news.
Most of this is older than the memes. People have always been fooled; corners have always been cut, the padded encyclopedia entry and the phoned-in review predate the transistor. Empty content had always existed; the economics are what changed. Producing the convincing-but-empty article used to cost roughly what producing the real one cost, which kept the supply scarce and the motive legible: somebody, somewhere, wanted something badly enough to fake it. Now the convincing version costs nothing and wants nothing. The scale is new, and the emptiness of intention behind it is newer still. For the first time the flood pours out of no one in particular, for no reason in particular, at volumes no manual forger could have dreamed. The old fakes at least had a faker, and a faker is a kind of stake: someone who chose, who could be found, who wanted the lie to work. That is the thing that has gone missing.
The feeds got the famous doses. Across 2023 and 2024, anyone who kept a social account watched the arrival of images that were almost wonders: a log-cabin interior with a waterfall where the fireplace should be; veterans with the wrong number of fingers holding signs asking why nobody says happy birthday; a crocheted Jesus emerging from the surf with a torso made of shrimp, ringed by comment sections saying Amen. Images like thousands of others that came from the same engagement farms, amplified by accounts that may or may not have contained people. When two researchers, Renée DiResta and Josh Goldstein at Georgetown, finally studied the phenomenon properly, they mapped a hundred and twenty Facebook pages running AI-image mills and found the platform’s recommendation systems actively promoting the output to people who had never asked for it, in volumes whose interactions ran to the hundreds of millions.[3] The images were doing something stranger than deceiving. They were succeeding as content while being nothing, collecting likes and shares and seconds of attention with no human intention anywhere behind them beyond the ad revenue of the page operator.
Music got quieter doses. By April of 2026, the streaming service Deezer was reporting that around forty-four percent of the tracks uploaded to it every day (some seventy-five thousand songs a day) were fully AI-generated.[4] The overwhelming share of their streams were fraudulent, bots performing music for bots while the royalty system kept score. A year earlier it had been a fraction of that. The synthetic flood had a human-run prequel: Spotify, as the journalist Liz Pelly documented in Harper’s,[5] had for years quietly commissioned cheap production music under invented artist names to fill its mood playlists (jazz for sleep, lo-fi for studying) on the theory that a listener drifting off does not check whether the pianist exists. The pianist, increasingly, does not.
Print got the doses that could hurt someone. Product reviews comparing products no one had touched, recipe preambles reminiscing about invented childhoods. In August 2023, the New York Mycological Society, the volunteer experts who have been teaching residents which fungi will and will not kill them since 1962, felt obliged to issue a public warning: the digital Amazon shelves were filling with machine-generated mushroom-foraging guides: handsome covers, confident identification plates, authors who did not exist, in the one genre where a wrong sentence poisons its reader; written by no one, checked by no one, sold next to the real thing with covers just as good.[6] The same month, author Jane Friedman found books for sale under her own name that she had never written: her byline, her niche, her plausible topics, but someone else’s machine output. Amazon initially declined to remove them because she had not trademarked herself.[7]
Whatever your particular knowledge diet (articles, feeds, playlists, search results, the morning scroll), all of it now draws from the same well.
The standard criticisms, deployed in a thousand columns and comment sections, are quality criticisms. The stuff is inaccurate, “it hallucinates.” The stuff is ugly, “look at the fingers.” The stuff is generic, “it has no taste.” Each criticism is true of some of it and getting falser by the next wave of model releases. But most do not touch the actual phenomenon, because the actual phenomenon includes the good stuff. The Glasgow images were perceived as beautiful. The fake foraging guides were fluent. The ghost pianists are, note for note, perfectly pleasant. If your account of slop is that it is detectable badness, then slop is a temporary problem better models solve or have solved, and the unease you feel should already be fading with those improvements.
Is it?
The account that holds up is narrower. Slop is technically adequate content produced by a process in which no one, at any point, staked anything real on its being right. Not wrong. Unstaked. Someone usually clicks publish, and often enough a name sits on the byline, but a name on the publish button still falls short of a person who would pay for being wrong, and that gap is what the reader eventually feels. The chain that normally connects a piece of content to someone who could be embarrassed by it has been broken, and what you are responding to, when something in the feed sits wrong with you in that low-grade, unnameable way, is that break itself: the sense that no one is on the other end. Content with no one behind it has a quality of nothing happening that is independent of its polish, and independent of its accuracy. You finish it and you have not met anyone. No mind committed itself in front of you. There was no wager, no matter how small. No one anywhere said I think this is true and worth your time, and I will pay something if I’m wrong, and a piece of content is valued by its wager. Strip the wager out and what remains is a void with excellent production values.
Before that account can do any work, it has to get past a louder set of objections than the aesthetic claim: the moral claim; the claim that fills the angrier side of the internet. “The machines are thieves,” trained on the work of people who never consented and will never be paid. “They are gluttons,” each generated image boiling off its sip of fresh water in a data center parked in someone else’s drought, drawing down a grid someone else has to power. “They are a desecration,” soulless engines aping what only a person has any business making. And under all of it, the oldest fear: “they are coming for us,” and the supporters cheering them on will be eaten last. Not one of these is the complaint of a crank. The theft is real and mostly unredressed; the cost and the power are real and worth counting; the dread is the sane reading of the trajectory.
But each of them takes aim at the machine (its training set, its utility bill, its missing soul, its trajectory) and judges the thing by its parentage, leaving aside what it is and whether anyone stands behind it. This kind of outrage is what people producing slop are content with hearing, because it is something they can always answer: buy the renewable credits, license the training data, pay a person to press the button, or keep the feed delightful enough that the question never comes up. But the actual injury, the one where no one in the chain is exposed when the thing fails, sits exactly where it always was. The disgust is real, aimed a few degrees wide of a much larger target, and walking it back onto the target is most of what this book is for.
Two reflexes are available here, and both dodge the real question. One rejects anything a machine touched on principle. The other forwards whatever the model produces because it reads fine. Most people live between them, using these tools daily and uneasy about them daily, with no framework for the unease. The question worth asking: is anyone staked on it, which still sorts everything, while is it AI sorts nothing now that the tells are vanishing. The stance worth holding lies past both refusal and surrender: use the machine freely wherever being wrong is cheap to catch or harmless, and insist on an answerable human wherever being wrong has a price.
But some of this stuff is great, and people say so with a grin. The cat in the pirate coat commanding a ship through a storm. The orangutan filmed in flawless documentary style doing the week’s grocery run. People love these clips, sharing them by the millions, feeling no betrayal whatsoever; and they are right to feel none. A known fake that promises nothing but delight, and delivers it, has broken no contract. It is a cartoon by other means, and cartoons are an honorable genre as old as drawing. This is “Good Slop”: content whose goal is a momentary flicker of pleasure and which claims nothing about reality. Enjoy them freely; this book stays well clear of any temperance lecture about them.
The trouble begins one step over. The pirate cat announces its unreality; the grocery-shopping orangutan, rendered well enough, does not. By 2025, the most shared clips of this kind were trailed by comment sections evenly split between people in on the joke and others asking where the orangutan lives. The same engines, the same feeds, the same polish produce the clip that winks and the clip that doesn’t, the confessed fiction and the unconfessed, the candy and the thing shaped exactly like medicine. A food supply does not have to be mostly poison to be dangerous, it only has to make poison indistinguishable from dinner, and charge nothing for either.
We are already seeing this boundary dissolve in our most intimate spaces. Take AI-managed breakups. By early 2026, several apps offered to “handle the friction” of ending a relationship: you provide the context, the model drafts the text, and a synthetic voice can even deliver the call, hitting exactly the right notes of firm-but-kind closure. It is technically perfect; it avoids the stuttering, the recursive apologies, the raw human mess. But it is a betrayal of presence. To be broken up with by a machine is to be told that your shared history wasn’t worth the cost of a conversation. The “Good Slop” of a pirate cat delights because it claims nothing; the “Bad Slop” of an automated breakup repels because it mimics the form of a human claim while withdrawing the only thing that makes it real: a person who was willing to be there for the hard part.
And here is the fact at the bottom of it: you cannot reliably see the difference. Blind tests keep humiliating us on this point. Readers asked to distinguish machine prose from human prose perform near chance; in some published studies they prefer the poems of the machine. If slop were a texture in the words themselves, the tests would not keep coming out this way. The texture lives in the relationship between the words and the world, in whether the promise the content silently makes is backed by anyone, and relationships do not show up in a sample. The parents of Glasgow could not have detected the fraud by studying the pictures harder. The pictures contained no fraud. The fraud was in what the pictures were attached to.
For most of a century the Turing test sat on the horizon as the field’s distant finish line: a machine whose conversation no one could reliably tell from a person’s. It was always decades away. Then it was a few years away. In 2024 the machines were close but still short: interrogators picked GPT-4 as the human a little more than half the time, below the rate at which they correctly picked the actual humans. A year later the same experiment, run on GPT-4.5, put the model ahead of that line. The blind tests are what crossing it looks like, and the line a generation of computer scientists had walked toward was passed so quietly that the culture’s main response was to coin a pig-trough word for the output.[8]
The model passed for human, and that passing is why passing-for-human can no longer be anyone’s test. Turing’s bar asked whether the words could fool a person. The question it left out is whether a single person stands behind them, and that question, the one the famous test was designed to make unnecessary, is the only one left that still sorts anything.
Glasgow, in particular, was catchable. In February 2024 the machines still left fingerprints: a pasadise of sweet teats sat on the website for anyone who read past the pictures, and a wary parent had something real to find. But the tell only ever marked the era’s immaturity, not the slop itself, and the immaturity is on a quarterly release schedule. The misspellings are gone now, the hands have the right number of fingers, the light falls the way light falls. Each new model wipes another fingerprint, while the thing the fingerprints used to betray, the absence of anyone behind the promise, is exactly as absent as before. Looking harder worked for a while. The window is closing as a matter of engineering roadmap, and the Glasgow parents’ mistake is on its way to becoming indistinguishable from due diligence.
And the vanishing fingerprints are the smaller half of the story. For a while the consoling thought was that the machine, even when you could not catch it, was only matching us, a very good copy, an A-minus forgery of competence. That thought is expiring too. On a widening list of specific tasks the output has stopped drawing level and started pulling ahead: the radiology read that catches what the tired resident missed, the translation that moves more gracefully than the staff translator’s, the explanation of a diagnosis at three in the morning that arrives clearer and more patient than the one the exhausted oncologist had minutes to give. Beaten outright, in the open, on the merits.
This is what unsettles the people who were braced for mediocrity and got something better. If the machine were merely worse, stakes would be an ordinary quality-control story: keep a human on hand for the hard cases. But the machine is increasingly better at the very things quality used to mean, and that raises a sharper question: when the output is superior, what is a person still for? The short answer is the one thing superior output cannot supply on its own: someone who pays if it is wrong.
So the detection happens, when it happens, at the moment the attachment is tested. You find out the content was slop the way Dublin found out: by showing up, by cooking the recipe, eating the mushroom, citing the case law. In June 2023 two New York lawyers were sanctioned in federal court for filing a brief studded with cases an AI tool had simply invented.[9] The judge’s order is a portrait of men discovering, in the worst available venue, what their research was attached to. That was when the phenomenon was still novel enough to make the front page. By the middle of 2026 it was a genre: a lawyer and data scientist named Damien Charlotin was keeping a public tally of court filings caught citing cases that an AI had hallucinated, and it surpassed fifteen hundred, climbing by five or six a day.[10] That June, a federal judge in Mississippi sanctioned the lawyers on both sides of a case and cancelled the trial.[11] An Oregon judge handed down a $110,000 penalty he called “a notorious outlier in both degree and volume.”[12] A veteran federal prosecutor was fired over a brief he let a machine rewrite and did not check.[13]
Slop is invisible in consumption and unmistakable in collision. Which explains the otherwise baffling shape of public feeling about it: billions scroll it contentedly all day (idle attention is exactly the consumption mode in which nothing is staked on the content being real) while the word slop itself spreads with the fury of people who keep getting burned at the specific moments they actually leaned on what they read. The feed does not feel like a crisis. The collision always does.
It also explains who felt the wrongness first and hardest: the people for whom a piece of content is never just a way to pass time, because their work obliges them to lean on what it claims. The doctor reading the literature leans on every sentence of a study. The judge leans on every citation in a brief. For people in that position, fluent unstaked content was never an aesthetic nuisance; it was a contamination of the supply they draw from professionally.
In February 2023, the science-fiction magazine Clarkesworld did something no editor does lightly: it closed submissions. Its founder, Neil Clarke, had watched machine-written stories arrive faster than his small staff could read them, hundreds in a month, then hundreds in a week, and the reason he gave came down to arithmetic, not quality. Some of the fakes were perfectly competent. A magazine of that kind survives on the attention of editors who read everything that comes in, and that attention is finite in exactly the way machine submission is not. The flood did not have to fool anyone to win. It only had to cost more to sort than the sorting was worth, and the supply of it has no bottom.
Teachers met the same flood from the other side, and teachers are supposed to have the antidote: the practiced nose, the instinct that whispers no one wrote this. Press the teachers who actually grade the papers and the instinct turns out to be something else. What they catch lives outside the prose, which is clean, frequently cleaner than the student’s own. What they catch is a discontinuity: the kid who spent September unable to hold a paragraph together forming a different, November dialectic; the voice on the page that lands a register away from the voice in the third row. A recognition in the person they already know, and the place where the person stops and something else begins.
The proof that the detection lives in the relationship and not in the text is what a good teacher does next: she cannot grade a suspicion, so she asks the student to talk. Explain this argument. Defend this claim. Show me how you got here. The answer comes back in a room, in a body, under questioning, the one place a text can never be taken. By then she has stopped grading the essay and started asking a person to account for it.
The institutional response confirms the diagnosis from another angle. Schools bought detectors, and the detectors became a scandal of their own: scores treated as verdicts, an arms race lost a model generation at a time, and real students flagged for the crime of writing plainly (one Stanford study found the leading detectors flag the majority of essays by actual non-native English speakers[14]). The tool the schools are converging on instead is software that watches the document rather than the text: the edit history, the keystrokes, the two thousand words that arrived in a single paste at 1:14 a.m. The software has given up on judging the words and tries to certify the making instead, provenance for homework, a chain of custody for the essay, which is the surveillance industry conceding that the question still answerable about a piece of writing has stopped being is this good or even is this human, and become what is this attached to, and how did it get here.
So the teacher’s famous instinct: she was reading the attachment, not the page, testing a bond she was in a position to test, weekly, in person. Detection by texture always failed, and detection by software is losing on schedule. Detection by relationship, a known person watched over time and asked to account for herself face to face, is the only kind with a record of working. Schools are meeting this first, because they collide with the flood on a grading deadline. The instrument everyone keeps waiting for stays imaginary; what exists is a person who knows you, in a room, asking you to say it to her face.
Embarrassment keeps surfacing in these descriptions. Confronted with any piece of content, in any medium, ask one question: is there anyone who would be embarrassed if this turned out to be wrong, or empty, or a betrayal of the person consuming it? Not legally liable. Not annoyed at lost revenue. Embarrassed; the social cost paid by someone whose judgment was on display and shown deficient.
If the answer is yes, you are holding the work of someone with stakes, however flawed the work. If the answer is no, if you trace the chain from the content backward and arrive at a process, a pipeline, a purposeless page operator, you are holding slop, however flawed or flawless.
Run the test on the cases. The mushroom guide written by a mycologist with her name on the cover, who knows that one wrong plate in a field guide ends her standing among the people whose respect built her career; that book has stakes, and its errors, if it has them, have an address. The machine-extruded guide uploaded under a pen name by an operator running two hundred titles: slop, even on the pages where it happens to be right, and especially there, because its correctness is uninsured, indistinguishable from its failures until the test arrives with dinner. The student essay drafted with a model, then rebuilt, argued with, signed by someone prepared to defend every claim at the seminar table: not slop. The same essay submitted unread at 3 a.m.: slop, and notice that in this last comparison the two texts might be word-for-word identical. The test does not examine the text. It examines what the text is attached to.
There is a relief within the test: you are not meant to run it on everything. Almost nothing that scrolls past you asks anything of you, and a life spent auditing every sentence for who stands behind it would be a misery and a waste. The test is for the few moments in a day when you are about to lean on what you read: the coder about to ship the model’s snippet into something that matters, the father checking whether the dose is safe for a child, the pastor about to preach a quotation as fact, the student deciding to trust the summary instead of reading the book, the buyer about to wire the deposit. At those moments the question earns its keep: behind this, is there anyone who would be embarrassed if it were wrong? The rest of the time, let it wash past. This skill is a small ask: a feel for which moments are the leaning ones, not the expertise to judge all the rest.
It is an odd place to have arrived: a definition of a content problem that cannot be applied by looking at content. It offends the instinct that says quality must live in the artifact, assessable by anyone with taste and no other information. That instinct was always mistaken. What we have been calling quality always spread beyond the artifact; a third of it always lived in the attachment between artifact and maker, and we could afford to ignore this for five hundred years only because the attachment came bundled automatically. Producing competent work used to be so hard that its existence proved a committed human. Competence was the receipt. The machines did not destroy quality. They destroyed the receipt.
What that did to the price of everything is the next chapter. But the human stakes of it are already visible in a detail from the Glasgow warehouse that the meme cycle mostly missed. The actors stayed. Knowing the production was a travesty, handed gibberish to perform for crying children at jelly-bean rations, several of them kept improvising through the afternoon (one, in the now-famous photographs, kept performing in a half-empty room in a magician’s coat) because there were children in front of them, and the children were real even if nothing else was. The images on the website had no dignity because they had no capacity to be answerable for the disappointment they caused. But the actor in the silver mask, disappointed as they may have been at the absurdity, had dignity, the capacity to be answerable. They stayed in the room when the collision arrived. The images, by definition, could only disappear.[15]
That pattern recurs throughout this book, and it is worth stating once as a rule of the era we have entered: the cost of unstaked content never disappears. It is transferred to whoever shows up.
Notes (15)
February 24, 2024. CBS News and others. ↩︎
Irish Times and RTÉ, October 31, 2024. ↩︎
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. ↩︎
Deezer reported that about 44 percent of tracks uploaded to it daily (roughly 75,000 a day) were fully AI-generated as of April 2026, the great majority of their streams fraudulent. Deezer Newsroom and Music Business Worldwide, April 2026. ↩︎
Liz Pelly, “The Ghosts in the Machine,” Harper’s Magazine, January 2025. ↩︎
August 2023. New York Mycological Society notice “AI Generated Mushroom Books: Life and Death” (Aug. 2023); 404 Media, “AI-Generated Mushroom Foraging Books Are All Over Amazon” (Aug. 29, 2023). ↩︎
August 2023. CBC Radio; Gizmodo. ↩︎
In a controlled study (Cameron R. Jones and Benjamin K. Bergen, UC San Diego, 2024), interrogators judged GPT-4 to be the human 54 percent of the time, below the 67 percent rate at which they correctly identified the actual humans; on that measure GPT-4 did not clearly pass. A 2025 follow-up using GPT-4.5 put the model at 73 percent, above the human rate, the first robust pass. “People cannot distinguish GPT-4 from a human in a Turing test,” arXiv:2405.08007. ↩︎
Mata v. Avianca, Inc. (S.D.N.Y. 2023). Attorneys Steven Schwartz and Peter LoDuca were sanctioned $5,000 by Judge P. Kevin Castel in June 2023 for a brief citing six cases that ChatGPT had invented. ↩︎
Damien Charlotin (HEC Paris). damiencharlotin.com/hallucinations. ↩︎
Judge Sharion Aycock (U.S. District Court, N.D. Miss.), June 9, 2026. Mississippi Free Press; 404 Media. ↩︎
U.S. Magistrate Judge Mark D. Clarke. ABA Journal, 2026. ↩︎
Assistant U.S. Attorney Rudy Renfer, 2026. ABA Journal; NBC News. ↩︎
Weixin Liang, James Zou, et al., “GPT detectors are biased against non-native English writers,” Patterns 4, no. 7 (July 10, 2023). ↩︎
The performers came out of it better than the production did. Paul Connell, who played the event’s Wonka, turned the day into a stand-up show and a wave of press, and the families themselves later organized a free, properly run event for the children. The clearest case is the teenage actress who improvised the Unknown: the role’s notoriety led to her being invited to train and perform at the London Dungeon, where she appeared in April 2024. The unstaked images caused the disappointment and answered for none of it; the people who showed up in the room were the ones who carried both the cost and, here, what little upside there was. STV News; Clydebank Post. ↩︎