Books in the Age of the Machine: A Retreat to the High Ground

Americans bought 782.7 million print books last year. That’s up 21% on 2019. E-books peaked back in 2014 and have gone flat since. So before we hold the funeral, look at the corpse: it keeps buying itself hardbacks.

And look at what’s doing the lifting. Not textbooks, not business manuals. Romance print sales roughly doubled between 2020 and 2023 — from about 18 million copies a year to 39 million. Bolt fantasy onto the romance and you get “romantasy,” which took seven of last year’s ten best-selling titles. Manga more than doubled. Most of it caught fire on BookTok, where teenagers talk each other into novels; those titles have grown five years running. Notice the pattern: none of the growth is information. It’s fiction, feeling, and people telling other people what moved them. Hold that thought — we’ll come back to it.

Let me reframe the question.

What people really want to know is whether books go the way of radio after TV, cinema after streaming, or newspapers after Facebook. It’s the right question, because those three endings are not the same ending.

Radio didn’t die. It still reaches about 80% of Americans every week, and among people under fifty it now beats broadcast TV — 83% to 59%. It lost the crown and kept the country. Cinema is shaky — the 2024 box office came in about three billion dollars light — but the same people streaming on the couch still show up at the theater; roughly three in four went at least once in the last six months. Radio and cinema each backed into the one thing they do that nothing else does, and lived.

Newspapers are the scary one. US print ad money fell from $49.4 billion in 2005 to $8.8 billion in 2020. Newsroom jobs dropped from about 74,000 in 2006 to under 31,000 by 2020. The paper’s still on the porch; the thing that paid a reporter to chase one story for six months is gone.

So which one is the book?

Here’s the catch, and I think it’s the whole game. The newspaper collapse was never about the medium. Nobody voted that reporting was worthless. The money left — the ad model that paid for real journalism walked over to Facebook and Google, and the journalism starved behind it. Books run on their own quiet money: the advance, the editor who spends a year knocking an argument into shape, the reviews that tell you what’s worth ten hours of your life. AI doesn’t need to make anyone hate books. It just has to wear down the reader’s patience the whole thing stands on.

But look at what AI is really good at, and what it isn’t.

It’s brilliant at looking things up. Ask it anything and it hands you the answer, tidy and instant. So a book whose job was to hold facts — the encyclopedia, the dictionary, the manual, the travel guide, the textbook that’s four-fifths facts — is already a museum piece. Wikipedia started that funeral. AI shuts the lid.

What lives is the book that builds a way of thinking instead of storing stuff. Kuhn’s The Structure of Scientific Revolutions matters not for its facts but for the shape of its argument. You can’t look your way to that shape. Someone has to build it and put their name on it.

And that’s the other thing AI can’t fake: putting your name on it. A book by a named human carries something no smooth machine paragraph does — this person has a stake, a reputation, a life standing behind the claim. When the world is flooded with decent text, more words are cheap. Someone answerable for them is not.

Now the part hardly anyone says out loud.

AI learns from what people wrote. If we stop writing the real thing and let the machines churn out the next round of “content,” the models end up feeding on their own leftovers. There’s a name for what happens next — model collapse — and it’s not a figure of speech. A 2024 paper in Nature showed that models trained over and over on their own output rot: they lose the rare cases and slide into grey mush. Say it plainly: serious human books are the seed corn of AI itself. Burn the seed corn for a fatter harvest this year and next year you’ve got nothing to plant.

But keeping the humans writing is easier said than done.

Researchers have even pinned down the recipe: models stay healthy as long as fresh human writing keeps getting poured into the training mix — you accumulate it, you don’t let synthetic text replace it. Fine. But that just hands the problem back to us. Why would people keep doing the slow, hard work of real books if AI has quietly eaten both their readership and the money that used to come with it? This is the newspaper story again, every gear in the same place: the writer does the work, the platform harvests the attention, and the trade that made the work thins out and goes cheap. AI needs the human corpus to stay fresh, and is busy draining the reason anyone would produce it.

There’s one lever, and it’s already being pulled — quietly. The AI companies have started paying for the corpus. News Corp took more than $250 million over five years from OpenAI; smaller outlets get offers of one to five million a year. It’s the Australian move writ large — the bargaining code that pried around $140 million a year out of Google and Facebook to fund news. Call it a market for “keep the humans writing,” being born in real time.

Two catches, though, and they’re big. It’s fragile: Meta simply stopped renewing its Australian deals in 2024 and walked away from the table. And the money lands on the publisher, not the person — the authors whose books sit in those training sets mostly weren’t asked, can’t opt out, and see none of it. So we’ve built a way to pay the company that owns the words, and still no way to pay the human who spent three years finding them. How do you reward slow, uncertain, years-long work in a world that can remix the result for free by morning? Nobody has cracked that. It may be the heaviest unanswered question the whole AI age hands us — and no, I don’t have the answer either.

So is anything actually safe?

Fiction, funny enough, is the fortress — not because a machine can’t write a novel (it can, and it’ll improve) but because a novel is a long stay inside one particular human head. Nobody reads Beloved for information. They read it to spend time with Toni Morrison. Same reason a flawless studio track never killed the live show. And remember who’s buying: the whole print rebound is being driven by exactly this — readers chasing a voice, not a fact.

Poetry might even gain. It’s the form where every word is one person’s exact choice, made because they felt something. Take the person out and the deal falls apart. A machine sonnet is a card trick. A human one is a confession.

Children’s books are safe for a reason that has nothing to do with words. The book is a prop for a relationship — a parent, a lap, a turned page, a small finger jabbing at the picture. I’ve done that bedtime a thousand times, and I promise you no app is invited. No model replaces it, and no model wants to.

And the deep reporting that comes from someone spending three years in an archive or a war zone keeps its worth exactly because that knowledge isn’t online yet. AI can only mix what’s already been digitized. The person in the room is still the only way the room reaches the page.

There’s a quieter loss in here. Writing a book is a kind of forced march — it makes you live with one idea for years, rewrite it, argue with yourself, and be wrong where everyone can see. That grind grows a kind of thinking a quick AI-helped draft never will. Kill the reason to write books and you don’t only lose the books. You lose the sort of mind that writing one builds.

Which leaves the real shortage. Not information — we’re drowning in that. Curation. Books were never just containers; they were chosen containers, someone’s judgment about what deserved your full attention. AI grows the pile of knowledge faster than ever and also turns it into a dump you can’t walk through. Someone, or something, has to say: this is worth ten hours. That job doesn’t vanish. It becomes everything.

My hunch — and it’s only a hunch, because nobody sees the future, me included — is that books end up where opera did. Not dead. Loved. A money-loser for most who make them, held up by patrons and institutions, and still, now and then, reaching a depth no faster format can. Vinyl didn’t come back to beat streaming. It came back to be what streaming can’t.

That’s not the death of the book. It’s a climb to higher ground.

The open question is whether we hold that ground on purpose — or let the seed corn burn because this year’s harvest looked so cheap.