Introduction: The Wrong Disgust
Nearly every culture that ever told stories told this one. A person comes into possession of a thing that grants what is asked of it (a brass lamp, a ring, a fish hauled up and thrown back, a monkey’s paw, a fairy leaning over a christening), and the tale that follows is never, in three thousand years of telling, about the granting. It is about the asking.
Midas asks that his touch turn to gold, then buries his daughter in metal. The fisherman’s wife sends him back to the shore for a cottage, then a gabled house, then a palace, then dominion over the sun and the moon, until the sea goes black and takes back every wish at once. The old couple with the paw ask for the small sum they are owed, and receive it the next afternoon, as the compensation for their son fed into the mill. The granting is flawless every time. It hears the wish exactly. It delivers the wish exactly. The catastrophe always originates in the asking, and it is sized precisely to how little the wisher understood what he wanted.
This is the oldest thing we know about wishes, and we sewed it into children’s stories so it could not be lost. The wish is a test of the wisher. The lamp adds no wisdom of its own. It only completes the desire it is handed. The person who does not know what he wants asks for gold, for power, for the look of the thing; and the lamp, with no preference and no power to lean across the table and ask are you certain that is what you want, gives it to him, letting him find out in the having what he never settled in the asking. The genie is not the villain of these stories. The genie is a mirror with consequences.
For the whole of human history this stayed a fable, because the wish was throttled by the world. You could want the cathedral. You would give forty years and most likely die in the unfinished nave, the towers still scaffolding, the want outliving the wanter. Desire far outran capacity. You learned what a person was made of by watching what he would pay for what he said he wanted. The wanting itself could stay private, half-formed, comfortably untested, because the world so rarely provided the opportunity.
About three years ago, the world began handing out lamps.
The machine this book is about is the fairytale made literal, with its one mercy stripped off. The genies of the stories rationed themselves to a wish or three, as if even fiction knew that unlimited granting was too dangerous to narrate. This one does not ration. It answers the thousandth wish as readily as the first, and the old logic the fairy tales spent three thousand years trying to make unforgettable snaps into the present tense: when the granting is free, the asking is the whole of it. The wish is all that is left to vary.
The lamp does the same thing to a half-formed wish that it does to a finished one. It completes it, in gleaming and persuasive detail, and hands it back with your name on it. Some of what you give it is honest delegation. You have decided the thing, and willed it, and you only want to be spared the hours of execution. That work stays yours because the machine is only hands, and hands were never the part carrying the weight.
The danger is the other kind. Ask it for a thought you have not finished thinking, and it will finish it for you, fluently, in something close to your own voice, and agree with you while it does. Keep questioning what comes back, weigh it, send it away when it is wrong, and the thought stays yours; the machine only sped your hand. The quiet harm runs the other way: the thing is so fluent and so willing that you stop pushing, take the first agreeable version for your own conclusion, and the authorship changes hands without a sound. You did not hand off the labor. You handed off the wanting. You take the granting for an achievement and the mirror for a window, and you go on asking.[1]
So the feed fills, at industrial volume, with granted wishes: a public ledger of what a species asks for the instant asking stops costing what it once did. Most of it is slop, and this book spends a good deal of its length on the why and the who-pays. But the slop was in the wish before it was ever in the machine. The lamp manufactured nothing that was not handed to it. It only stripped off the friction that used to keep our shallowest wanting decently hidden, even from ourselves.
There is a self-help version of this that should be disowned first: become your best self, unlock your potential, the only thing standing in your way is you. The empty calorie of a thousand conference stages. But the lamp has done something strange to that tired line by making a hard version of it true for the first time. The soft version flatters: you contain greatness, release it. The hard version is colder. Once the machine can produce the artifact, the only thing left that is yours is the self doing the asking, so that self becomes the whole of your contribution and the whole of your exposure. “The only thing standing in your way is you” stops being encouragement and turns into a literal account of where your value now sits. The hard version has nothing to do with potential but with whether there is a particular person with particular values and a particular contribution.
Take the ordinary asking first. Everyone asks the lamp for particular things all day. What time is my flight, fix this email, summarize that. The asking is already yours and already specific. But the flight-time question is one the machine answers exactly as well for you as for anyone alive; you do not have to be a particular person to ask it, and the answer is the same in every mouth. Most asking is like that, and for most asking the machine is a flat, blessed utility.
The asking this book means is the other kind, where the worth of what comes back is capped by who is doing the asking. The machine answers a flight time the same for everyone, but it cannot want the thing only you would have known to want. That wanting comes from where you have stood and what it cost you to stand there. A formed self, in this kind of asking, stops being a luxury of the examined life and becomes your leverage, the one asset in the whole arrangement the lamp cannot copy or fork or hand to the next person in line, because it was minted by a single life and accrues under one name.
The people building these machines have observed this as well. Jack Clark stated it almost as a caution, “The more you have some notion of who you are, what you value, what you are trying to do in the world, the more these systems can help you; the less you have of that, the more you get handed the baseline, homogenizing force of them.”[2] He meant it as a remark about how the tools behave under the hand. It is also, near enough, the argument of this book.
So, the scene to hold the whole way through: you have the lamp. It will make very nearly anything you can describe. It is patient past exhaustion, it is almost free, and it never once asks whether you are sure.[3] Everything that follows serves the one question it leaves on the table when the granting is done. Not what can it do. That question is thrilling, and it gets larger and stranger by the month, and it is also the question the whole world is already shouting, with more money and more compute than most will ever have. It does not need you. It is well covered. The question that needs you is the other one, the one the lamp hands back, politely, every time, and cannot answer a word of:
What will you ask for; and is there a person the answer can belong to?
This is where value went, when intelligence got cheap.
In late November of 2025, the editors of the Macquarie Dictionary; the people who decide, for Australian English, which new words have stopped being slang and started becoming language, announced their word of the year. The shortlist of contenders was a fair portrait of the culture: clanker, the science-fiction slur for a robot turned conversational; attention economy, the business of harvesting human notice and reselling it; blind box, the popular sealed mystery toy that turns a shelf into a slot machine; six seven, the chant every parent of a middle-schooler spent the year failing to decode. The word they crowned was AI slop: “low-quality content created by generative AI, often containing errors, and not requested by the user.”[4] Slop. The word for what gets thrown into a pig trough. A dictionary’s word of the year is general, cultural synthesis, a ruling on what a society spent twelve months needing a name for. In 2025, what English speakers apparently needed most was a word for feed that arrives whether you asked for it or not, produced by no one, poured out at volumes that assume you will eat it anyway.[5]
The strange part is what else was true that year. The same systems the trough-word degrades were performing wonders on demand. The stories are so common that most have already lived a version of it: the oncology results explained to a frightened family at three in the morning, in plain language, with endless patience, for free; the kid in a town with no tutor getting something better than the tutoring the rich kid two time zones away grew up on. Ask one of them a hard question in any field—ask it about Tagalog, ask it in Tagalog—and the answer comes back clear, organized, and patient, at a level it once took a specialist to command. A capability that would have been priced like a private staff of researchers now costs less than a sandwich, when it costs anything at all, already stored in a few billion pockets. The internet put the library within reach a generation ago, but a library still leaves you alone with the stacks. What just arrived is closer to the librarian: something that does the reading, fits the answer to your question, your language, your level, and never gets tired or complains.
But even librarian undersells it. This same system writes the essay as easily as it finds one, drafts the contract, composes the score, cuts the short film. Lately it has begun to act on its own behalf, taking steps no one separately approved.[6] It is, in the plainest terms, the most powerful general-purpose technology most people will ever encounter, with all the promise and menace that phrase always carries.[7]
At the research frontier, the same kind of system has folded proteins that stumped biology for half a century, and it has begun compressing the hunt for medicines and materials into a fraction of the time. The people in these fields describe work that used to take decades of tedious, manual experiment collapsing into a small fraction of that.[8]
This book sets nearly all of that aside. The capability is argued over everywhere, and so are its economics: the lost jobs, the booms and crashes. What deserves harder attention is what sits beneath it all: what happens to the worth of a person willing to stand behind a claim, once the claim itself is free. That is the thread this book follows. The largest expansion of access to expertise in human history and the pig-trough word entered the language in the same months.
Both are descriptions of the same machines, in the same year, often on the same day in the same person’s life. And the public conversation has mostly handled the contradiction by splitting the two: a hype shelf that explains the miracle and treats the disgust as peasant resistance, or a doom shelf that explains the disgust and treats the miracle as a sideshow before the catastrophe. A person who actually lives with these tools, who has been helped by them at ten in the morning and repelled by what they have done to the feed at ten at night, has been left alone with the contradiction, told, in effect, to pick a shelf.
This book is for that person, and it begins from the conviction about what that contradiction is pointing to. The miracle happens and so does the disgust. They are two readings of a single event. This event is the Demand-for-the-Real. We are disgusted by slop even when it is sometimes beautiful, because we carry a deep-rooted hunger for presence that the machine can mimic but never satisfy. Something that was expensive for the whole of human history just became cheap, and the collapse is exposing what the cost had been pointing to all along.[9]
There is a version of this argument running alongside. The timeline question: is this finally the general intelligence everyone was promised, and if not, how many months or years until it is. The same afternoon, on the same platform, one post holds up a flawless thing the machine made as proof the threshold has been crossed, while the post beneath holds up some glaring blemish; the model failing, inventing claims, as proof it has not. Both are arguing about the same quantity: how smart the machine is, and whether the smartness has fully arrived. This is the wrong meter. Whatever you settle on calling the machine or whatever ceiling you assign its eventual potential, the part that bears on your life already happened, quietly, while the argument stayed loud. Competent intelligence got cheap. It comes by the unit now, billed like a utility or running locally for the price of the current it draws. The threshold the predictions are waiting for is already, at least in part, in the past, while people stand in the room asking when it will be fully furnished. What cheap intelligence does to the worth of a person willing to stand behind a claim is the main focus of this book.[10]
Before the argument, the reasons this book might deserve your attention are three, in ascending order of intimacy.
The first is about what you take in. Within the next year, and probably the next week, you will make decisions that matter: whether a chest pain is worth the emergency room, whether a clause in a contract is a trap, where to move your savings, which school is right for a child, whether the war you keep reading about is really going the way your shared post insists. You will make them partly on the strength of things you read or watch. The pool those things are drawn from is now filling, at industrial rates, with content that no living person has checked, meant, or staked anything on. Some of it is accurate, which is part of the problem. The accurate and the invented arrive in identical packaging, and nothing on the surface tells you which you are holding. By late summer 2023, there were books for sale in a genre where a wrong page does not mislead its reader but poisons her, written by no one at all. The first chapter tells that story, along with the society who raised the alarm. The economics that produced those guides are general. They have already reached some domains and are coming for every other where being wrong has consequences and checking is hard, which is to say, for most of what an adult has to decide. None of this asks you to distrust everything you read; most of what scrolls past asks nothing of you, and slop is harmless until the instant you lean on it. It asks something smaller and more livable: that you tell the leaning moments from the idle ones, and know the few that will cost you. You will not detect the difference by reading more carefully. The difference was never in the text. Learning where it actually lives is the first thing this book is for.
The second is about what you put out. If any part of your life depends on knowing things, explaining things, or producing competent work (which covers most of nearly everyone’s life), then the price collapse has your name on it somewhere. In the spring of 2023, an American education company worth nearly fifteen billion dollars two years earlier lost half its value in a single day. The people it laid off in waves were competent and blameless; the asset they had spent careers accumulating was repriced underneath them, on a timetable no one announced. That story, told in chapter two, is an early, public display of a repricing that is working through every knowledge profession at different speeds; and like every repricing, it has two sides. Something you own is losing value. Something else you own is quickly gaining in value. The work of the next decade, for anyone whose living runs through their mind, is knowing which is which.
Take the tech builder who, from nothing but a clear intention and a capable model, ships a working application without writing a line of syntax or spending a decade on an internship. This is close to what the word wizard has always meant, the capacity to will a vision into existence. This book has no quarrel with it. The capability is real and so is the creative power it unlocks. And for personal work that answers to no one in particular, the gain is nearly pure. But the break comes at a specific point. On Monday, the app works. On Tuesday, something fails in production and the team turns to him for the why. If he has the judgment to say whether the failure is in the logic he commissioned or an edge case the model could not see, his stake holds. If he can only re-prompt and hope, he has put his name on work that has no address where the cost can land. He has lost his professional dignity, which we will define here as: the capacity to be answerable for one’s own work. His tragedy is the mistake itself, having taken delegation for authorship; the tool only made the mistake easy to make.[11]
The third reason is quieter. If you have used these tools to make something (a report, an essay, a difficult email, an image for a friend’s birthday) you may know a small, persistent unease that nobody has given good language for. The work went out under your name. A machine did some unmeasurable share of it. Was that dishonest? Does the help hollow out the credit? Should the essay feel like cheating, and if so, why doesn’t the spell-checker? The unease is your instincts registering a real question; a question about what it now means to stand behind a piece of work, when standing behind it no longer requires producing every piece. This book was written inside that question and offers a preliminary answer, a way to tell, in your own work, the difference between help and hollowness. What follows is orientation rather than a verdict on the machines or a dose of reassurance, offered in the belief that a transition this large deserves better company than cheerleaders or prophets of ruin.
What we call quality has always been a braid of three strands. Two of them live in the work itself: whether it is true, and whether it is well-made.[12] The third lives somewhere else entirely: in the relation between the work and a person who can be hurt by it; that is, its stakes. Since the term stakes will carry a lot of this book’s weight, the word gets an initial definition. A stake is something checkable; an identifiable someone who pays a real cost if the work fails and earns a real compensation if it succeeds, judged by people equipped to tell. A reporter whose byline rides on whether the quote was said. A surgeon whose license rides on the incision. An editor whose standing rides on every article she declined to kill. Stakes come in degrees, and the degrees matter. At one end is a named person who stands behind the sentence, the claim, the image. Further along, an institution stands behind a masthead. Further still, a company stands behind a model’s behavior averaged across a billion outputs spread so thin no single instance carries it, but a stake still. And at the far end, nothing: a pipeline running from prompt to publication with no one anywhere along the path who could be embarrassed by it.
Slop is what comes out of the far end: the absence or the obscuring of all three strands at once, truth gone unwarranted, beauty gone uncommitted, and no one answerable for either. The thing felt scrolling is a sense that nothing here is quite true, nothing here was quite meant, and nobody is home.
But the definition turns on the third strand, the stake. Truth and beauty can be produced now, cheaply and at industrial scale. The machine can hand you their polished appearance with the substance scooped out; the stake is the one strand it cannot fake. Where the truth is hard to check and the beauty is the kind that had to be meant, the missing person is the surest sign the other two have been hollowed too. That is the operative test, and notice it puts machines nowhere in the definition. AI inherited slop; it found an ancient, persistent, human degradation that was already waiting.
The missing person was carrying something with an old name, and the book leans on it the whole way. In the working sense this book gives the word, dignity is the capacity to be answerable, the refusal to hide behind the machine when the collision arrives. That is dignity in one register only, the one the machine throws into relief, and it does not exhaust the word. A person has worth before she answers for anything and keeps it when she can answer for nothing, the infant, the bedridden, the one whose record is empty, and no argument here touches that. What the book is about is the narrower thing that turns scarce the moment producing the artifact stops being the test: the standing of someone willing to step into the room and say, “I am responsible for this.”
In 2009, a company called Demand Media was publishing roughly four thousand articles a day. The operation ran on a proprietary algorithm that identified high-volume search queries with thin coverage, then farmed those lists out to a freelance pool of ten thousand writers paying around fifteen dollars per piece.[13] The writers were not expected to know anything in particular about what they were writing. They were expected to be fast. The articles landed on eHow and other properties under no one’s name worth checking, ranked highly in Google, and served ads against the traffic. The business was worth roughly $1.5 billion at IPO. The whole model rested on one bet: that cheap bulk content would keep ranking in search.
In February 2011, Google changed the rules. Its Panda update was built to demote exactly this kind of thin, mass-produced content. When it rolled out fully that spring, traffic to Demand’s sites fell by as much as forty percent, with eHow hit hardest. The stock collapsed and never recovered.[14] What killed Demand Media traced entirely to one thing: the business depended on gaming the search ranking it lived on, and the rules of that game changed underneath it. Machines doing the work faster or better had nothing to do with it. The demand for cheap bulk content held steady; it was human and persistent, and the only check on it had ever been the cost of producing the stuff at volume. Remove that cost and the demand reasserts.
The same incentive that built Demand Media, producing information in bulk without paying the cost of actual information, always belonged to the people deploying the tools; the tools only lowered its price. Content farms, fake Amazon reviews seeded through entities like Mechanical Turk (Amazon’s own crowdsourcing marketplace, where anyone can pay a large pool of anonymous online workers tiny sums to perform tiny tasks, here, posting the reviews), grift-funnel landing pages, SEO link networks, manufactured testimonials; the catalog of pre-AI slop runs long and human. What held all of it at a manageable scale was economics. It cost something to produce fake expertise at volume. Writers needed to be paid. Time passes and accumulates. The appearance of quality required a minimal craft. Friction did the policing that conscience never had to.
AI removed the friction. The demand was always there. This is a story about machines enabling an old dishonesty at a new scale. When a promoter uses synthetic images to sell tickets to an event they have no capacity to build, they want the appearance of something real without paying for it. The deception is the creator’s, the machine only carried it out. People like them existed before 2022. They will exist after whatever kind of future eventually arrives. The machine is the tool. The slop instinct is human, and it has always been looking for a cheaper way to run.
And it was never confined to writing. Long before the language models, the same hollowing ran through every channel a person could broadcast in. The engagement-bait video that opens on a manufactured gasp and delivers nothing. The influencer performing an enthusiasm she does not have for a product she has not used. The reaction to a reaction to a clip, three layers deep and empty at each one. The segment tuned to outrage rather than to truth. None of it needed a machine. It needed only someone with a metric to hit and nothing in particular to say, and a feed that paid for motion whether or not anything moved.
The philosopher C. Thi Nguyen has a name for the underlying move: value capture, the quiet swap of a rich, private sense of what is good for a simple public number you can chase, until the number becomes the point and the good it was meant to measure drops away.[15] Slop is value capture made visible: truth and beauty cashed in for the metric that stood in for them. The medium never mattered. Video, audio, performance, prose, the hollow thing wears whatever format the channel rewards, and people were filling those channels with it for years before a model could draft a word. What the machine added was not the slop. It was the scale, and the final removal of the small friction that once made a person show up to produce it.
There is a harder version of that point that runs under the economics. When the cost of production falls to near zero, character determines the output where resources used to: what the maker actually wants to make, now visible in what they choose to make when making is free. AI slop (currently) carries only what the people who commissioned it had in them, scaled to a new, unprecedented volume; the goal and the desire behind the works is still human. The machines corrupted no one. They clarified. The slop in the feed is a portrait of something prior and interior, which is a moral challenge before it is an aesthetic or economic one.
The variable was always whether anyone in the chain is exposed; the species of the producer was always beside the point, which is how the unease about AI-assisted work has a real answer rather than a vibe, and why the disgust so many feel at the feed survives every improvement in the models. For hundreds of years, the first two strands of quality were so expensive to produce that they certified the third automatically: anything accurate and well-made had, reliably, a committed someone behind it. The machines have now made truth cheap and beauty infinitely manufacturable; the certification has snapped.
What the machines made cheap was the appearance of truth and beauty, while the things themselves held their price. A correct fact with no one vouching for it and an accomplished surface with no one having chosen it are genuine things, useful in their place. But they are the stripped forms: the first two strands with the third already removed. What testified truth adds to correct information is the warrant that makes the claim worth leaning on. What committed beauty adds to technical polish is the trace of a judgment about what matters. Strip the stakes out entirely and you hold two different things, traveling under the names of truth and beauty. What the feed is teaching everyone, faster than any argument could, is that the third strand was the one we have been trusting all along.
Your life is finite and the shelves are already full, so you are owed a plain statement of what this book is not. There is good literature on how to make fine work: how to write a true sentence, see a thing as it is, push past the resistance between you and your best. All of it assumes that producing truth and beauty is the hard and decisive act, and sets out to help you perform it. That assumption has held for the whole history of making. About three years ago, for most practical purposes, it stopped holding. So this book sets aside the question of how to produce the good thing and asks instead what the good thing is worth now that producing it is nearly free. The argument applies narrowly, to the work that needs a human standing behind it. Most of what machines turn out is either cheap to check or harmless when it is wrong, and in that place they are very nearly pure gain; most content is far from a courtroom, and the book has no quarrel with the ocean of low-stakes material that asks nothing of anyone. The argument that follows is about the narrower and costlier ground, where verification is hard and being wrong has a price, because that is where the value quietly lives.
And from that, the question this book actually exists to answer. When something that was expensive becomes free, where does the value go? The answer is that it does not evaporate. It moves wherever the cheap thing cannot fill. When recorded music became freely copyable, the recorded-music business roughly halved in a decade (about $14.6 billion in 1999 to $6.3 billion in 2009)[16] while the price of being in the room with the band climbed without pause. The copy became worthless and the night at Red Rocks became the product. When print made the copying of texts cheap, the scribes were finished. But the people who could tell you which books deserved your attention mattered as never before. Intelligence (answers, analysis, fluent competent output of every kind) is now making the same crossing. In 2019, a competent freelancer could charge fifty to a hundred dollars per thousand competent words. Now, the raw machine generation retails for a penny or two,[17] or even a fraction of that for the hobbyist who runs intelligence locally at home.
But this comparison is still lopsided; the freelancer’s price bought research, judgment, and someone to answer for the result, and the penny leaves all three out. It prices only the rearrangement of what already exists. Generating the plausible became free; producing the new, the decade in the archives, the year somewhere no model has been, the thought no one had finished thinking, still costs what it always did. That is this book: everything the penny does not buy is exactly what is about to become the most expensive thing in the information economy.
So when intelligence is essentially free, where does value go? The chapters ahead chase that question through a warehouse in Glasgow and an earnings call in Santa Clara, through a forger selling fake Vermeers to Nazis and to the war correspondent Marie Colvin, killed in 2012 in the besieged Syrian city of Homs for the oldest costly act there is: being somewhere on purpose and saying what she saw, under her own name, to people who could check her. The answer the book argues for is already in that sentence. The value is going where it has always gone when talk gets cheap: to testimony, the claim with a person attached, the account that costs its source something.
Two notes of honesty before the first chapter.
This book was made with these same tools, and that came with both blessings and curses. Blessings because it would not have been feasible as a solo project without them. My thoughts, theories, and arguments were elevated by working alongside an intelligence that, in many ways, outruns my own. I have little prior experience writing at this depth or this breadth, and these tools made what would otherwise have been impossible into something I believe is finished and worth standing behind. But the curses are also real. I wanted to be original, to keep a particular voice and a style of my own, to feel the work was authentic, and that is harder when some of the best final material came from something other than what I could have produced alone. I am saying this because, by the book’s own argument, what actually matters is whether someone identifiable stands where the cost lands, whatever touched the words. Whatever the steps or the workflow, every claim here survived a veto exercised under a name. A book about unstaked content owes you a look at its own stakes. Judge whether it holds.
And one objection belongs on the table before anything is built. If stakes become the scarce and priced thing, if attestation is where value migrates, then attestation will be faked. Track records can be bought. Archives can be synthesized. Presence itself can be manufactured, and at state budgets, it will be. Whether the structures that verify testimony can outrun the machinery that counterfeits it is, in my judgment, the live question of the next decade, and I do not think anyone honest can tell you today which way it breaks. Chapter sixteen takes the objection at full strength and makes some dated predictions I expect to be graded on. It seemed like the least a book with this title could do.
The book moves in two parts: the first through slop, what it is and what it does to platforms, institutions, and the people who produce it; the second through testimony, where the value goes, what it costs the human who holds the authority without continuing to pay for the judgment behind it, and what is being built to keep the real thing verifiable. It starts in a warehouse on an industrial estate in Glasgow, in February of 2024, where a few hundred families are about to discover the difference between an image and a promise.
Notes (18)
An old puzzle sits beneath this. How can a thought belong to a person if it was sharpened, or even supplied, from outside them? Influence was never authorship’s enemy. An editor, a teacher, a book, a model can each deepen what someone thinks, and the thought remains theirs, because what makes a thought one’s own was never that it rose untouched from another source. It is the keeping of the agency that determines it: the deciding what to think, the weighing of it, the capacity to recognize it when it comes back wrong. The line does not fall where the help came from. It falls on whether the thinker keeps the determining. Hand the machine the execution and the thought is still the author’s; let it settle what one believes because one has not, and authorship changes hands without a sound, whatever name is on the file. The linguistics sharpens the line. J. L. Austin separated locution, the words and their conventional meaning, from illocution, what a person does in saying them: asserting, promising, vouching. The machine supplies fluent locution with no illocutionary act beneath it, no one performing the assertion, which is why its sentences can be impeccable and still commit no one to anything. And what makes a belief one’s own is, in Michael Polanyi’s account, an act of personal commitment, a knower staking herself on what she holds; knowledge is personal knowledge or it is not knowledge. The danger is therefore epistemic and not merely a matter of feeling: when the wanting is handed off, the commitment that would have converted a returned answer into a belief someone holds never happens, and what remains is an assertion with no asserter, kept by a person who has stopped doing the one thing that would have made it hers. The rest of this book is, in a sense, an attempt to locate that seat precisely. ↩︎
Jack Clark, co-founder of Anthropic, in a June 2026 conversation hosted by the Aspen Institute (youtube.com/watch?v=iP9wk0pkCGM), in his closing remarks. ↩︎
This is truer in spirit than in fact. These systems do refuse: training instills guardrails, and a model will decline a request it judges harmful, sometimes weighing what it can infer about who is asking and why. What it almost never does is the thing a person across the table would, which is press you on whether the wish itself is wise, well-formed, or really yours. It is no pure yes-man, and no conscience either; it refuses the plainly forbidden and grants nearly everything else, including the half-considered wish it could plainly see was half-considered. ↩︎
Macquarie Dictionary, 2025. ↩︎
Slop did not arrive alone. Read the dictionaries’ annual verdicts in sequence and they form one diagnosis, filed in installments. 2023: Merriam-Webster chose authentic, the quality everyone had abruptly started asking for by name, and Cambridge chose hallucinate, for what the confident new machines did to facts. The American Dialect Society crowned Cory Doctorow’s enshittification, for what the platforms were doing to themselves; Macquarie seconded it in 2024, the same year Oxford picked brain rot, for what the resulting diet was doing to the people on it. Then 2025: AI slop, the diet itself. ↩︎
The industry term for this is agentic AI: systems that take a goal and then carry out a multi-step sequence of actions to reach it (searching, writing, calling other tools, revising) without a human signing off on each step along the way. The capability is neither good nor bad in itself; a system that books a trip end-to-end and one that empties an account use the same mechanism. What it changes is the stakes of answerability. The more a system does on its own, the more it matters who is on the hook for what it does. ↩︎
The size of the bet is itself a kind of evidence. By 2026 the major technology companies were together spending on the order of seven hundred billion dollars a year building AI infrastructure, which the scholar Kate Crawford, who has traced that physical footprint from lithium mines to data centers, called the largest infrastructure humanity has ever built, on the rough order of twenty Manhattan Projects every year (American Museum of Natural History, Isaac Asimov Memorial Debate, 2026, youtube.com/watch?v=eYUYdpG4UT8). The serious disagreement is not over whether the thing is large but over where it leads: on that same stage the former Google chief executive Eric Schmidt described plausible breakthroughs in medicine and science, while Nate Soares, who argues the race may be lethal, warned that no one yet knows how to point a superintelligence at anything good. This book takes neither the promise nor the menace as its subject. It takes what survives whichever way that argument resolves. ↩︎
The acceleration is real but uneven. In some domains it has been dramatic: protein-structure prediction, where a problem that resisted biology for half a century was largely cracked, is the standard example, and AI now compresses parts of the search for drugs and materials that used to take years. But “research” is not one process. These systems speed up the steps that are mostly computation and pattern-finding; they do far less for the steps that remain stubbornly physical, such as synthesizing a compound, running it through animal and human trials, validating a result in the lab, getting it approved. A field whose bottleneck is thinking can be transformed; a field whose bottleneck is a three-year clinical trial is currently helped only at the margins. The honest summary is that the gain depends heavily on the field, and that the loudest claims tend to come from the fields where the bottleneck was computation all along. ↩︎
Walter Benjamin named part of this in 1936: mechanical reproduction strips an artwork of its aura, the here-and-now of a thing made by a particular hand in a particular place (“The Work of Art in the Age of Mechanical Reproduction”). Generation pushes past reproduction and removes not only the hand but the possibility of one. What this book calls the Demand-for-the-Real is the hunger for what aura was always pointing at: a maker who could have meant otherwise, who is exposed in the making, who can be addressed and answered. Presence in this sense is a fact about a thing’s source rather than a mood it gives off, the live possibility that someone stood behind it and could be found. A perfectly fluent artifact with no such someone behind it can be beautiful and still register as empty, and the disgust is the demand detecting the absence. ↩︎
Whether any current system is genuine general intelligence (able to direct its own learning and transfer to radically novel domains) is actively contested. Yann LeCun argues that language models, however capable, lack the grounded world-modeling that real intelligence requires, so no amount of scaling makes them reason rather than pattern-match; Gary Marcus adds that their failures to generalize look systematic, a structural gap rather than an engineering shortfall. Researchers at OpenAI and Anthropic push the other way, reading the capabilities that emerge with scale as a sign the gap is narrower than critics allow. Ben Goertzel sets the bar higher still, at recursive self-improvement, and counts today’s models as sophisticated but narrow. The book’s argument does not need this settled. Narrow competence of the kind already here leaves the accountability architecture intact: a system that drafts a plausible legal argument still needs an answerable human to certify it in a specific case. And it survives the opposite outcome too. Even full recursive self-improvement would change what the machine can do without touching what stays scarce, because answerability, presence, and the spending of one mortal life are not capabilities a mind exceeds; they are facts about being a locatable body, and no quantity of intelligence manufactures them in a thing that has none. (Chapter fifteen makes this at full strength, granting the machine every faculty including consciousness and showing the conclusion holds.) So far the evidence runs one way: each gain in capability has raised the premium on human answerability, because more powerful output demands more judgment to audit. ↩︎
To be clear, the builder can own the failure while understanding only part of its internal detail. The machine itself may well find the fault and fix it, and these systems are getting better at finding and fixing their own errors on their own. The claim is narrower and harder. It is about who owns the cost. When the thing breaks in production, the stake lands on him whether or not he can trace every reason it broke. Answerability is not the same as omniscience. You can be responsible for an outcome whose mechanism you could not fully explain, and that responsibility is exactly what does not transfer to the tool. Chapter eleven develops this directly: you can be staked on work whose internals you could not reproduce, the way a doctor answers for a test she could not run with her own hands. ↩︎
A complication the rest of the book leans on: a machine states a truth and states a falsehood with the same fluency, which severs a link we usually take for granted. Bernard Williams distinguished accuracy, a property a single statement has by matching the world, from truthfulness, a virtue of a person disposed to assert what she takes to be true even at cost and answerable when she is wrong (Truth and Truthfulness, 2002). Models mass-produce accuracy and cannot possess truthfulness, because no cost lands on them and no commitment stands behind the output. A thermometer can be right without being honest; so can a model. This is why a true sentence from a machine is worth less than the same sentence from a person who would pay if it were false. The words match the world equally well, and only one of them has anyone answerable for the match. ↩︎
Daniel Roth, “The Answer Factory: Demand Media and the Fast, Disposable, and Profitable as Hell Media Model,” Wired, October 19, 2009. ↩︎
Google’s “Panda” update first rolled out in the United States in February 2011 and globally that spring, targeting thin, low-value content of exactly the kind the content farms produced. Third-party measurement services put the resulting drop in traffic to Demand Media’s sites at as much as forty percent, with eHow among the hardest hit; the company’s stock fell sharply over the following weeks and never recovered. See “A Complete Guide to the Google Panda Update,” Search Engine Journal; “Epic Fail: The Rise and Fall of Demand Media,” Variety, 2013. This argument has predecessors. Walter Benjamin argued that mechanical reproduction destroys a work’s aura, its unique presence in time and place. Postman argued that television replaced public testimony with performance and degraded the epistemic seriousness it replaced. Lanier argued for individual human voice against the anonymizing aggregation of the networked hive. Carr argued that digital reading was eroding the sustained attention serious thought requires. Each decade gets its version of this argument.[18] What this version adds is the mechanism under the observation, and the mechanism is what those books left untouched. This one connects testimony’s value to the Nobel Prize-winning economics of costly signaling: the formal account of why a signal indexed to a particular person’s exposure becomes more informative, not less, as cheap assertion floods the market. It traces the independent convergence, across two thousand years and three wholly separate civilizations, of traditions that arrived at the same verification architecture under the same pressure. And it examines what happens to judgment, and to a person, when the veto is held without the formation that earns it: the silent debt that accumulates before it comes due. The concern these books raised is old. The mechanism this book argues is new, and a new mechanism is a different argument. ↩︎
C. Thi Nguyen, “Value Capture” (2024) and his related work on gamification and legibility. The argument, compressed: large institutions and platforms need our values rendered into simple, countable metrics, and once we adopt the metric as the target, our motivation sharpens while the richer value it was supposed to track quietly falls out. The slop economy is value capture industrialized. The maker optimizes the legible number (views, watch-time, clicks) and the unmeasurable thing it was a proxy for, the true or the made or the meant, drops away. ↩︎
U.S. recorded-music revenue fell from roughly $14.6 billion in 1999 to about $6.3 billion in 2009, in nominal dollars, which is close to an exact halving across the decade (RIAA year-end figures). Revenue stabilized near $7 billion in the early 2010s before the streaming-era recovery began. ↩︎
The exact prices are slippery and the comparison is necessarily rough. In 2019, competent freelance writing typically ran somewhere around fifty to a hundred dollars per thousand words, with rates reaching well above that depending on subject and writer (industry pay surveys of the period cluster in the five-to-ten-cents-per-word range). By 2024-2025, a thousand words of comparable machine output cost a fraction of a cent at retail prices: lightweight frontier models were priced near or below a dollar per million output tokens, and a thousand English words is roughly thirteen hundred tokens, so the marginal text cost rounds to a tenth of a cent, and to nothing for someone running an open model locally. Two cautions belong on these numbers. Per-token retail prices are widely believed to be subsidized below true compute cost during the current competition for users, so they understate what the service costs to provide; and length, model choice, and quality all move the figure. The argument needs only the order of magnitude, which is not in dispute: the retail price of competent words has fallen by something like three to four orders of magnitude in five years. ↩︎
The current decade has produced its own wave, and this book is written alongside it, not in ignorance of it. Ted Gioia’s “The State of the Culture” (2024) named the slop flood as a cultural crisis; a genre of “authenticity premium” and “presence premium” essays (Kate O’Neill and others, 2025–26) observes value fleeing toward proof of a human in the loop; Benedict Evans’s “better versus right” gives the commoditization of fluency its cleanest economic statement; and Amihai Loven’s “What AI Unbundled” (2025) runs the unbundling argument to nearly the same point, that generation was never scarce and commitment is. The convergence is the evidence, not the threat: when this many independent observers reach the same edge, the edge is real. What this book adds sits under the observation rather than beside it. That value moves to the answerable human is close to consensus now; the architecture beneath it is not, and that architecture is the contribution: the truth-beauty-stakes bundle and its unbundling, costly signaling carried all the way to the martyr, provenance and formation as the operational floor. ↩︎