AI Can Save Literature, and Publishing Can't

Literary culture withers if writing well does not lead to being read, but the relationship between a text’s quality and its discoverability is the weakest it has ever been. Self-publishers face inordinate marketing burdens and tech-platform algorithms that reward a pace incompatible with serious work. Traditional publishing is corrupt and exclusionary, and it has been allergic to serious literary ambition for at least thirty years. The processes by which excellent writers and readers are supposed to find each other are inoperative, and no one seems to know how to fix them.

We can’t expect this damage to heal on its own. Publishers already know whom they want to publish; they’re not going to start reading unsolicited manuscripts to find out if they’re wrong. Literary agents toss 99+% of everything unread because they know how hard it is to sell an acquisitions committee a book whose author hasn’t thirst-farmed their way to 300,000+ followers on Instagram. Curation quality is low because there’s no incentive to improve it; the industry cares more about the whims of publicists and chain booksellers than the reader’s or author’s experience. The book industry discovered enshittification long before it had a name.

I’m one of the world’s leading experts on the question of whether AI can read or write literary fiction. (Read yes, write no.) The economic incentives that drive AI progress favor technological applications rather than literary aesthetics—an agent that can debug a massive codebase is more useful than one that can imitate beautiful prose that doesn’t sell—and the paradoxical combinations of human traits necessary to write at the highest levels are impossible to induce in devices that simply mimic us. Although trade publishing will turn at least one AI-generated novel into a bestseller by the end of 2027—I’d bet it already has—AI can’t write the stories we talk about for decades after they’re written. Traditional publishing loves to blame the devaluation of the written word on AI, but it’s had a decades-long head start, and done most of the work on its own.

Authors know, with the pain and intimacy of a hemorrhoid, how difficult it is to get other people—especially the self-arrogated tastemakers who get paid to do it—to read. We’ve covered literary agents; it only gets worse from there. Ten to fifteen say-so’s sit on a typical acquisition committee, and only one or two of them will have read more than five pages of the book on whose fate they decide. A freelance copyeditor is often the first person to engage with the text at a level beyond a glance. Everyone’s too busy to read. Can AI help?

Controlled experiments have found flaws in LLMs’ reading abilities, but also that a signal can be extracted from modern frontier models, if the setting is properly designed, that is a superior indicator of textual quality than selection by a literary agent or traditional publisher. LLMs easily tell the difference between unpublishable and publishable work. They can sometimes separate, from a pool of mostly average traditionally published stories and essays, those publishable at elite levels—the lead titles, the column busters. They cannot tell the difference between an excellent-but-not-outstanding novel and our era’s Ulysses. That’s fine; AI doesn’t need to do that to be useful. At what traditional publishers claim to do as curators—wade through the slush to discover the few pieces worth reading—technology now outdoes them. If publishers’ claim of curatorial supremacy rests on subjective taste, then their tasteless behavior—the celebrity memoirs, the shoddy editing, the query system—invalidates their entire argument. On the other hand, if it relies on objective craft, we should expect a sufficiently advanced language model to be able to detect it.

I won't claim that AI will have only positive contributions to literature. Already, this is not the case. AI slop books have stolen attention, sales, and reputability from legitimate self-publishers. Undisclosed AI influencers, who unlike human ones never tire of the grind, will easily defeat their carbon-based competition; wealthy individuals and traditional publishers will use them profusely. AI could save literature, but I do not prophesy that it will do so, because I’m not a prophet. I speak of capabilities, not certainties.

Self-publishers live and die on technology platforms’ algorithms, but so do traditional publishers. These ranking systems were designed in the 2000s and 2010s when artificial intelligence was primitive and full-text analysis, unaffordable. They rely on social proof and coarse-grained quality estimates, both of which are easy for slop artists and grifters to game, because nothing better existed when they were made.

Consider, in a better world, an online bookstore whose recommendation system, unlike Amazon’s, uses no social-proof signals because full-text reading gives it all the information it needs. Obscure titles are discovered and promoted on the basis of literary excellence. Readers’ niche interests can be fulfilled. Reading time is not restricted to those who have the leverage or social status to get pages read, because everything is read. Nothing is skimmed dismissively and discarded because its protagonist has the same name as an intern’s better-looking or more successful college roommate. Deserving literary work finds an audience even if written by the sort of person—say she’s a 58-year-old schoolteacher in Nebraska with no literary agent and no family publicists through whom to find an agent—the current system is designed to ignore. The best way to “game the system” is to write well. This doesn’t exist now, and we all suffer for it.

What has traditional publishing given us in the past few decades? The query system. That’s its one real innovation. I don’t care to speculate on which dysfunctions in publishing come from bad faith—insiders would have a harder time publishing their friends if trad-pub darlings had to compete with real writers; no one seriously doubts that—and which ones come from the system being overwhelmed; either way, trade is a dead end. We need AI to save literature because we’ve tried everything else and nothing has worked.

If I’m wrong and the robot overlords turn us into paperclips, at least we’ll never have to read another one of trade publishing’s throuple memoirs.