When Knowing Got Cheap

On the afternoon of May 1, 2023, Dan Rosensweig got on his company’s quarterly earnings call and did something chief executives almost never do: he named the thing. His company was Chegg, and if the name means nothing to you, it meant everything to American college students, who had turned it into a verb. Chegg sold answers. A student stuck on problem 3 of a calculus set at midnight paid $14.95 a month, typed the problem in, and got the worked solution, drawn from a database of tens of millions of them, with freelance subject-matter experts, many in India, producing new ones around the clock for whatever the database lacked. Professors loathed it; students subscribed by the millions; investors loved what professors loathed. At its February 2021 peak the company was worth nearly fifteen billion dollars, priced over $110 a share. By the spring of 2023 it had already fallen hard from that peak, to around $17, as pandemic-era growth cooled and the market re-rated the stock over two years; the earnings call is the moment the slide turned into a cliff. Wall Street analysts liked to describe its “moat” as two decades of accumulated, proprietary, searchable knowing, more than five million paying subscribers deep.^[1]

The quarter’s numbers were fine. The guidance was not, and Rosensweig explained why with a candor that cost him a billion dollars before lunch the next day. Since March, he told the analysts, the company had seen a significant spike in student interest in ChatGPT, and now believed it was hurting the rate at which new customers signed up. That was the whole disclosure. Students had found a box that answered homework questions instantly, on any subject, for free or nearly so, and were declining to pay $14.95 for the older box. By the close of trading on May 2, Chegg’s stock had fallen 48 percent in a single session, to $9.08, from the roughly $17 it had traded at before the call. The financial press, reaching for precedent, found none; analysts started calling the pattern “getting Chegged,” and every software company’s earnings call since has included a ritual paragraph explaining why it could not happen to them; a paragraph that, three years on, is still there every quarter, as permanent a fixture as the revenue slide. It kept happening. It is still happening. Chegg cut four percent of its staff that summer, then twenty-three percent the following June, then more, in tranches that eventually totaled most of the company.

As of the middle of 2026, the stock trades for about a dollar (99 percent of the peak value gone) and the company has spent its decline laying off staff in deepening rounds while suing Google, on the theory that AI-generated answers at the top of search results were stealing from the little traffic it had left. The seller of answers, reduced to litigating against a machine that gives them away for free. The verb has already faded from campus. The students simply moved to the box, the way their parents once moved from the encyclopedia to the search bar, without ceremony and without looking back.

Notice what had been destroyed. The demand for answers survived; students wanted answers more than ever, and usage of the new tools overwhelmingly demonstrates that. What collapsed was the scarcity of answers: the condition under which an archive of solutions could function as a moat, a subscription, a salary, a stock. Chegg’s misfortune was to be the purest publicly traded expression of a much older asset class: the value of knowing things. Its collapse was that asset class being repriced in public.

Almost everyone reading this owns some of that asset class. So the question is not what happened to one Silicon Valley company but what your knowing is worth now, and to answer it you have to watch where the value went, because it did not vanish. It moved.

---

To see the size of what broke, it helps to remember how strange the old price of knowledge was, and how recently everyone still paid it without thinking.

In 1990, the Encyclopaedia Britannica company sold about a hundred and twenty thousand multi-volume sets in the United States, at well over a thousand dollars a set. It was the company’s best year ever. The product was thirty-two volumes of settled human knowledge, sold door to door by a sales force trained to sit at kitchen tables and convert parental aspiration into installment plans. The pitch worked because the underlying scarcity was real: if your child needed to know the boiling point of antimony or the succession of the Plantagenets at nine on a school night, the alternatives were the library tomorrow or the encyclopedia now. The salesman in the suit was selling a heavy, multi-volume promise: that the world’s knowledge had been audited, edited, and bound.

Within one generation that sales pitch became incomprehensible. Wikipedia arrived in 2001, built by volunteer amateurs, which was supposed to be the punchline, except that study after study found its science and history articles holding their own weight against the professionals.^[2] In March 2012 Britannica announced the end of its print edition after two hundred and forty-four years; the final printing had sold about eight thousand sets.^[3] The repricing took roughly a decade, and throughout it turned on one fact about the new thing: that it was there, free and instant. Being better never entered into it. The commodity had arrived, and the salesman’s suit, the costume of the staked authority, was the first thing to be repriced.

What happened to Britannica’s encyclopaedia sets happened next to the people. Stack Overflow, founded in 2008, became the place where the world’s programmers asked and answered each other’s questions; a commons of working knowledge, sixty million questions and answers deep, sustained by a currency of reputation points that working coders printed on their résumés. It was, in its domain, what Britannica’s door-to-door salesmen had promised: every answer, in the house. Then the models trained on, among other things, Stack Overflow itself. The programmers discovered that the box would answer their exact question, about their exact code, without the wait and without the hazing ritual of being told the question was a duplicate. The site’s traffic began sliding in 2022 and then fell off a cliff; by the spring of 2023, outside analysts were measuring month-over-month drops in the double digits. The company disputed the steepest numbers at the time, claiming 2023 traffic fell only five percent below the prior year. Nobody disputes the cliff anymore, since the volume of new questions eventually fell to levels last seen when the site launched. In October 2023 Stack Overflow laid off twenty-eight percent of its staff. The following May, it signed a deal licensing its archive (the accumulated, freely given knowledge of its community) to OpenAI, for training. The commons sold itself to its successor.

By the end of 2025 the site was taking in fewer than four thousand new questions a month, a level it had not seen since 2009, the year after it launched, back when it was a scrappy upstart rather than the place programming ran on. The place still stands, the way a cathedral town stands after the trade routes move, its archive more consulted by machines in training than by the programmers it was built for. Reputation points, it turned out, were denominated in the old currency.^[4]

And the freelancers. In 2019, a competent generalist writer charged roughly five to ten cents a word: fifty to a hundred dollars for a thousand words of clean, researched copy, and a workmanlike living for the people producing it at scale. A machine now does the same thousand words for Pennie’s. The comparison is lopsided in the writer’s favor on everything except price: the freelancer’s fee bought research, judgment, and a person who could be called, embarrassed, or sued, none of which the penny covers. That was cold comfort on an invoice. The billable unit had always been word-production, with accountability thrown in free. Most working writers, on most days, were being paid because producing competent words was hard. It stopped being hard, and the line items adjusted within months: Duolingo cut about ten percent of its contractor translators at the end of 2023, with a spokesman conceding the link to the company’s adoption of large language models. Studies of the freelance platforms have also found writing and translation gigs measurably degrading almost immediately after ChatGPT’s release.^[5] Forecasts argue, severance memos do not.

Across these cases the shape is identical. For the entire prior history of the species, knowledge had a body attached. To get an answer, you compensated a human life that had spent years acquiring it. Someone bought the encyclopedia that paid the editors, someone paid the subscription that paid the experts. Hire the writer, ask the doctor, retain the lawyer at six minutes a billing increment. Every fact came bundled with a person, and the bundling set the price. The companies deploying language models completed a process the internet had started, and cut the body off the answer. Knowing-that, the recallable, articulable, explainable content of expertise, exists as a standing reservoir, dippable by anyone, priced like tap water, or owned outright when running locally on a machine for the cost of the electricity to spin it. Intelligence arrived first as a metered utility you buy by the sip and is becoming, in parallel, a household appliance you can own: either way a thing, severed from the person who used to come attached to it. The archaic correlation between knows things and worth something has broken, and it broke in roughly thirty-six months. Almost every institution built on the old correlation (schools, credentials, reference publishers, mid-level expertise of every salaried kind) is still standing in their current form because institutions take longer to fall than correlations.

Call that alarmism if you like. Whether it is alarmism or arithmetic turns on timing, on how long an institution can keep standing after its premise has gone. And timing is the question on which the nineteenth century ran the perfect experiment.

---

In 1806, Frederic Tudor began shipping pond ice from New England to the Caribbean. It was an absurdity of logistics: harvesting frozen water in Massachusetts, packing it in sawdust, and sailing it to places where the sun was a permanent adversary. He was jailed for debt and mocked in the papers, but he had identified a fundamental truth: cold, in a hot place, is worth almost anything. By the 1850s, he was the “Ice King,” presiding over a global industry that shipped Walden Pond’s surface as far as Calcutta.

What followed is the correlation displayed in practice. Mechanical refrigeration was invented within Tudor’s lifetime, yet the natural ice trade did not collapse overnight. It grew for decades more. Machine ice was at first expensive and unreliable; natural ice had the brand, the distribution, and the installed base of iceboxes. A man cutting ponds in 1890 could argue that the machines still hadn’t won since before he was born. Then, around the turn of the century, the machines crossed the threshold. The natural ice industry went from enormous to nonexistent in twenty years. The men who had read the long coexistence as refutation discovered it had been incubation.^[6]

The pattern matters for us because the coexistence phase is where knowledge work is right now. The doctor, the analyst, the staff writer, and the tutor coexist with the box, and the coexistence is being read as evidence the repricing has stalled. Chegg is what the threshold-crossing looks like from inside a single firm: years of nothing, and then minus forty-eight percent in a day. The lesson of the ice trade is that displacement is nonlinear. The only people who navigate it well are those who ask early: what, exactly, were people paying us for: the ice, or something the ice was carrying?

---

The collapses keep revealing the same thing: the dying price was almost always a partial price. Something else was bundled in, invisible until the bundle broke.

Recorded music made the cleanest demonstration. In 1999, the American recorded-music business took in $14.6 billion. A decade of perfect free copies later, it took in about half that. The albums had been repriced to zero. But over the same decade, North American concert revenue went from $1.7 billion to over $4 billion. The price of being in the room climbed as relentlessly as the falling price of the recording.^[7]

What the concert numbers describe is the liberation of a preference that had been there all along, suppressed by an older logic. When the album cost money, the night was the bonus. Once the recording cost nothing, the disguise came off. The people who had been buying the album had always been buying the Demand-for-the-Real: the presence, the irreproducibility, the fact that a specific room filled with specific air at a specific hour would never exist again. The recording had been a way of approximating that, a souvenir selling for the price of the thing it commemorated. Set the souvenir free and the real thing reasserts itself.

The same unmasking has been running across every domain where copy met physical presence. Vinyl record sales were declared dead in the 1990s, only to come roaring back in 2008. By 2023, vinyl was outselling CDs for the second year running.^[8] This is the unmasked demand for an object with weight, a ritual with steps, and the twelve inches of someone’s deliberate attention made permanent on a disc. They are buying the receipt the stream cannot supply: evidence that someone made this and it exists in the world as a thing that can run out. The farmers market runs the same proof in tomatoes: people pay more for the person who grew them than for the fruit itself.^[9] The person turns out to have been part of the value all along, hidden by the convenience of the supermarket.^[10]

Scarcity hid the demand for the real experience, the way the album hid the demand for the night, and kept it there all along. When the copy gets free, the demand surfaces, having been there since before anyone built a distribution system around it. The market the machines are creating is the original prize, depressed for two hundred years by the cost of reproduction, now visible because reproduction has hit the floor. The reader who pays for the signed edition, the listener at the hall, the patient who insists on the doctor who will answer if she is wrong; these are all the buyers who finally know what they were shopping for. And the epistemic version of this hunger is the want, now pricing into everything, for a person in a room who can be wrong about what they said and answer for it. The hunger is for testimony, the thing information was always carrying and could never replace, and testimony is rarer than the information that floods every screen.

Gutenberg ran the same experiment on the scribes, with the same result and a longer settling time. Printing repriced the copying of texts to near nothing, and the copyists were finished. But a Europe suddenly drowning in books needed something it had never needed at scale before: people who could tell you which books. Editors, commentators, indexers, the whole apparatus of learned judgment that the scarcity of manuscripts had made unnecessary. The flood created the filter. Reproduction collapsed; selection was born as a profession.^[11]

And then the newspapers: the bundle whose breaking cost the most, because what died in it was a kind of witness. American newspaper classified advertising peaked around $19.6 billion in 2000, roughly forty percent of the industry’s ad revenue: the most boring money in media, paying, invisibly, for the least boring asset in civic life. When Craigslist and its kin made listings free (researchers later attributed about five billion dollars of the displacement between 2000 and 2007 to Craigslist alone), total newspaper advertising fell from $49 billion to $26 billion inside the decade, and the thing the boring money had been subsidizing was exposed to open air. Used-car sellers, it turned out, had been funding the reporter at the Thursday-afternoon zoning meeting. That reporter had always been paid through someone else’s invoice. The classifieds had carried her the way the album had carried the band, and when the bundle broke, information got more abundant every year of the collapse while coverage shrank: the person who was there, who knew the aldermen by sight, who saw through the press release because she had sat through the meeting it was misrepresenting. We will spend a later chapter on what the loss of her turned out to cost. For now she is the third instance of the pattern: break a bundle, and you discover what was really being paid for, always too late to bill for it.^[12]

So run the pattern forward on knowing. Intelligence, in the form of answers, explanations, drafts, competence on demand, is the copyable half of a bundle we never knew was a bundle. The question that ought to keep every knowledge worker, every school, every publisher and every creator up at night is the other half: when the answer itself costs a penny, what was the body carrying that the penny does not buy?

You can hear the market already give an initial answer. The consoling story says the human simply climbs a rung: from answer-giver to judgment-giver, from coder to architect, from teacher of facts to mentor of persons. Some of that is real, and some of it is already true; the same months Chegg was collapsing, demand for in-person tutoring held and in places grew. But the rung you climb to is not safe either. The senior engineer who knows where the system breaks in production is worth more than the junior who only knew the syntax, at least until the models have ingested enough production to close even that gap, which they are visibly doing; ask the engineers watching an agent refactor a codebase how needed they feel this year versus last. Whatever can be stated as expertise (written down, demonstrated, made explicit) is what these systems are built to absorb, and “move up the value chain” is sound advice only until the chain is being eaten from below faster than anyone can climb. So the migration is narrower and stranger than “become the expert,” because expertise is copyable in principle and is being copied in fact. The value is draining toward the one thing in the bundle that stays put through all of it: the person who is answerable for the answer: who pays if it is wrong, who was in the room, whose name is on it. What attaches to them is stakes, the thing the bundle was always quietly charging for.

The pattern turns on its makers, too. The companies now selling intelligence as a metered service are not exempt from what happened to Chegg. They are exempt from it in practice, for now: the best frontier models still outrun what consumer hardware can run locally, the capability gap is real, and a meaningful premium tracks it. This is the coexistence phase. The ice trade had fifty years of it; recorded music had a decade. In each case, both sides kept advancing: the floor of what the cheaper alternative could do rose year by year, and so did the ceiling of what the premium version achieved, which is why the gap compresses more slowly than the impatient assume and the threshold arrives more suddenly than the incumbent expects. What the history makes clear is the shape, whatever the duration: the transition is threshold-dependent and nonlinear, and the companies that navigated it won on the strongest surrounding structure rather than the best commodity: the cold-chain relationships, the installed base, the things the commodity was carrying without appearing to. When ice got cheap, the companies that only had ice were done.^[13]

What distinguishes the frontier AI labs from one another, in this frame, is what they have built besides the benchmark. Several of them have started to talk in these terms: a lab’s published safety commitments, its stated accountability standards, the research it releases as a form of self-imposed liability are described as the durable asset, the structure that outlasts the ice. Treat that story with the suspicion it has earned. A safety posture marketed as a moat is still, in part, marketing, a way of attaching a premium to outputs that the next cheap local model may match anyway, and the firm selling it has every incentive to call its own conduct infrastructure. Whether any of it functions as cold-chain structure, or only as a story told to investors, is the thing commoditization will test and has not yet. The firms whose premium tracks most directly to raw intelligence quality, and least to anything surrounding it, are where Rosensweig was: selling access to the commodity in a market that is learning to manufacture the commodity cheap. The investors backing them appear to believe the benchmark is the moat. The ice merchants believed that too, for decades, and for a long time they were right. The question they will eventually answer, whether they mean to or not, is the one this chapter has been asking all along: what, exactly, was the bundle carrying?^[14]

It is worth saying plainly that the labs are not blameless weather in this story. The same companies now offering to sell their accountability as a feature are the ones that trained on Stack Overflow’s freely given commons and on a generation of writers’ work without asking, then sold the result back into the markets those people lived on. OpenAI bought Stack Overflow’s archive after the traffic that built it had already collapsed; the commons was monetized once by its members for reputation and a second time, over their heads, as training data. That is a choice some named firm made, in a boardroom, with a signature on it, and “the technology made knowing cheap” should not be allowed to launder it into something that merely happened. The repricing was built and shipped by people who can be named, and who profited.

Blame is one thing the repricing has earned; it is not the thing this book is chasing. Naming the other half of the bundle, the part the penny does not buy, is the work of the next several chapters. The usual names for it (judgment, taste, trust, expertise) each circle the thing without landing on it, because each still describes something a person knows how to do, and the machines are coming for everything a person knows how to do. The repricing runs past all of them. It moves value from what is said to who is answerable for it, from intelligence, which has become a commodity, toward something older that never can be, because it cannot be copied even in principle. Every previous migration in this chapter ended somewhere physical: a concert hall, a kitchen table, a council chamber. This one is headed the same way.

But the repricing of knowing has stirred up a fight that should be settled before the value is followed any further: the cry, rising from every corner of the internet where makers gather, that the machines did not reprice the sentences; they stole them. If knowing was never the asset, what exactly does anyone own in a sentence?

Notes (14)

1. Peak: $113.51/share (February 12, 2021), market value ~$14.7 billion. MacroTrends; CNBC, “Chegg drops more than 40% after saying ChatGPT is killing its business” (May 2, 2023). ↩︎

2. Jim Giles, “Internet encyclopaedias go head to head,” Nature vol. 438, pp. 900–901 (2005). Wikipedia launched January 15, 2001. ↩︎

3. Encyclopaedia Britannica, founded 1768. CBS News, “Encyclopaedia Britannica ends 244 years of print” (March 2012). ↩︎

4. Stack Overflow, founded 2008. VentureBeat (Oct. 17, 2023); SiliconANGLE (May 2024). ↩︎

5. TechCrunch, “Duolingo cut 10% of its contractor workforce as the company embraces AI” (Jan. 9, 2024). ↩︎

6. The natural-ice trade coexisted with mechanical refrigeration for roughly half a century before collapsing around 1900. “History of refrigeration.” ↩︎

7. Fanning’s program was Napster. RIAA year-end figures; Pollstar. ↩︎

8. U.S. vinyl LP sales surpassed CD sales for the second consecutive year in 2023 (some 43 million units), the first time vinyl outsold CDs since 1987, according to the RIAA year-end report for 2023. The demographic data consistently shows a majority of purchasers under 35. RIAA; Rolling Stone. ↩︎

9. The number of farmers markets registered with the USDA grew from roughly 1,755 in 1994 to more than 8,600 by 2019, with growth in suburban and urban areas outpacing rural ones; the buyers least in need of a local food source grew fastest. USDA Agricultural Marketing Service. ↩︎

10. This is not a claim that the farmer’s tomato tastes objectively better; sometimes it does, often it is indistinguishable, and the supermarket’s supply chain is frequently the more reliable one. The premium is paying for something else. The buyer is choosing a findable person who grew the thing and will answer for it over an anonymous corporation optimizing for margin, and is willing to pay for the difference even when the fruit itself is a wash. The preference is for an answerable source, which is the same thing this book keeps finding underneath every premium that survives a commodity going cheap. ↩︎

11. Gutenberg’s movable-type press (c. 1440) made copying cheap and displaced scribes (Parisian copyists attacked a press in 1476) while creating new demand for editors and curators of the resulting flood of books. ↩︎

12. The $26 billion figure is for 2010. Seamans and Zhu, Management Science 60(2), 2014. ↩︎

13. Ice itself never lost its value, and you can still buy a bag at any gas station. What died was harvested pond ice, the business of cutting frozen lakes and shipping the blocks, made obsolete the moment manufactured ice was cheaper and more reliable to produce. The commodity survived; the expensive way of producing it did not. That is the precise shape of every case in this chapter: the thing people wanted (cold, answers, music) persisted, while the costly method that used to be the only way to supply it was repriced to nothing. ↩︎

14. The comparison to the ice ship is deliberately imperfect, and worth being honest about, because a frontier lab is not only selling a commodity. Two things can keep a premium alive even after the cheap local model is “good enough” for ordinary use. First, raw frontier capability does work the floor cannot: novel drug and material discovery, research at the edge of what intelligence can do, problems where the difference between the best model and the second-best is the difference between a result and nothing. Second, convenience and reliability are themselves worth paying for; the best model on tap, maintained, current, and answerable when it fails, has value beyond the benchmark number. The chapter’s point survives both concessions: benchmarks alone are a weak moat, because the floor rises and the gap compresses. But “weak moat” is not “no premium,” and the labs that endure will be the ones whose worth is not reducible to a leaderboard. ↩︎