On AI-Generated Literature

If AI is not good enough to write literature right now, it will be soon. To convince myself of this, I had the Anthropic Claude model generate a “missing chapter” of Franz Kafka’s fragmentary novel Der Prozess that should fill in a gap between Chapter 9 Im Dom and 10 Ende. To give it the best chance of impressing me, I selected the most capable model variant, Opus 4.7 with adaptive thinking. I gave it a hyperlink to the book, but I did not give it any other instructions besides to write that one missing chapter. ⊕You can read the chapter for yourself in https://claude.ai/public/artifacts/e480a75e-9889-4e6b-9300-18d1b8e995ab Here is the opening it generated in English translation:

When K. returned from the cathedral he found that the gas lamps along the embankment had been lit, although it was scarcely past four in the afternoon; the fog that had filled the cathedral seemed to have followed him out into the streets, where it lay so thickly that the lamps appeared as small yellow blooms suspended in nothing, without any column or post to hold them up. He walked slowly, partly because of the fog and partly because he did not wish, just yet, to be in his apartment, where he would be obliged to think; so long as he was walking he was, he reasoned, doing something, even if that something amounted only to walking. The Italian, of course, would not be coming now. K. had waited an hour and a half at the cathedral, and the priest’s words had constituted, he supposed, a kind of explanation, though of what precisely they were the explanation he could not have said.

The chapter continues with K. drafting a petition to the court late into the night. In the end, he realizes the petition is futile:

He laid the pen down. It struck him that the petition could never be completed, not because of any failure on his part, but because the form of it required completion to be impossible: every event admitted further events, every justification required further justification, and no act of writing, however prolonged, could encompass a life, since the life continued throughout the writing, and indeed the writing was itself part of the life, and would have to be included in the petition, as would the inclusion of the writing in the petition, and so on, in a regress whose terms K. could now perceive with a clarity that surprised him.

There is an odd interaction with the Deputy Director of the bank, and then the chapter ends with the proceedings of the trial seemingly on hold.

As you may be able to appreciate, the result of this little experiment was not amazing. For every sentence that reads a bit like Kafka, there are a couple more that just don’t. I won’t belabor this point. I will just note that this opening is in fact an improvement over the opening to this same missing chapter generated with GPT-4 in 2024: ⊕For analysis, see Kafkaesque Algorithms: Kafka’s Writing in the Age of Artificial Intelligence by Jake Fraser, Humanities 2025, 14(1), 13; https://doi.org/10.3390/h14010013

After his discourse with the priest, Josef K. vacated the cathedral and roamed the city streets, his mind heavy with ruminations of the enigmatic court. The world around him seemed to twist and shift subtly, the familiar urban terrain taking on an uncanny, dreamlike quality. It was as if the very essence of the court had bled into the fabric of the world, transmuting it into an embodiment of the labyrinthine bureaucracy that had ensnared him.

Based in small part on this comparison, but in greater part on the rapid general advancement in this field, I do not hesitate to predict that AI will soon be capable of writing passable literature. Large language models are great at pattern matching, and text is a low-bandwidth channel, making it particularly amenable to pattern matching. ⊕This is why a literary critic wouldn’t have any of it if you told him that you discovered a previously unknown Sylvia Plath poem at an estate sale. You found it in some old guy’s sock drawer, and the verse is just so melancholy that it has to be hers, right? Of course, the critic would simply scoff. I concede that the literary critic might entertain your claim if you produced a much longer text, a fact which is consistent with the low-bandwidth of text as a communication channel. Nevertheless, handwriting analysis and a paper trail documenting the provenance of the document will be the deciding factors, not which exact words come in what exact order. On the other hand, an art historian might actually agree to take a look at that painting you found at a garage sale which you think might be a Grant Wood. Visual art is a higher-bandwidth channel, and accordingly harder to fake. Likewise, I would not be surprised to learn that AI can faithfully churn out full novels with Save the Cat or Hero’s Journey plot arcs. This is probably done by orchestrating LLM calls with an agentic workflow and cleverly managing context. Let’s not marvel at this ability. Let’s call AI’s apparent authorial adroitness a fine technical achievement and leave it at that.

The ability of AI to write books is already the premise of several pop-up generative AI products that can be surfaced by a simple Google search. One such product, AIWriteBook, claims to be able to take a book idea, generate a plot arc, crank out the text, and publish it on Amazon–all in just a couple hours. Its competitor, Squibler, lists testimonials from many satisfied customers. “Squibler knew where my story was going even when I didn’t,” says one author who apparently turned to AI after struggling to think up the plot for his novel.

But is anyone really reading AI-generated novels? At the time of writing, AIWriteBook claims that over 15,700 users have already used it to generate more than 58,000 books. These numbers sound inflated, but then again, the world is a large place. About 11,000 new titles get published every single day, so this would just be a drop in the pond.

Nothing has puzzled me more in the past week than the realization that people might actually be publishing and reading AI-generated books. Why do people read literature in the first place? What is the very best reason you can give for reading and re-reading your favorite book? Is it the plot? The author’s voice? Those are important, but if I am right those will soon be modeled well by AI–and besides, they don’t quite hit the nail on the head. If you dig deeper, I think you will find that the reason you keep re-reading War and Peace, The Lord of the Rings, or I Know Why the Caged Bird Sings is because there are a few scenes in there where the author reaches right through the page, and for a moment you share a seemingly psychic connection with him. You recognize yourself in a situation and the author reassures you he has been there before, that you aren’t totally alone in this world. That’s the mountaintop experience of reading your favorite book. Plot twists and drama are essential, but if literature weren’t capable of delivering that mountaintop experience, then you and I wouldn’t read. We would just watch YouTube videos or listen to podcasts or go golfing instead.

It is precisely because that mountaintop experience is the paramount pleasure of reading that the reader should insist a work of fiction be able to deliver it before he even picks it up. But an AI-generated story simply cannot deliver this mountaintop experience as I describe it. That is not because the AI cannot perfectly execute Save the Cat, hit that perfect balance between “show” and “tell” in dialogue, or diligently avoid point of view drift. Rather, there is no AI-generated scene that can make you realize that the author has “been there before” or that you aren’t totally alone in this world–for the very fact that the author does not exist. That brute fact is inherently alienating, and there is no sugar-coating it. If you had a mountaintop experience reading some AI-generated book, you would have to grapple with the fact that the putative “author” also had that exact same experience, only he had it in that moment when ChatGPT.com first streamed those inspired words of his novel onto his internet browser. At risk of absurdity, we must consider as dashed the highest hope a reader might have when cracking an AI-generated book. And if that hope is dashed from the outset, what reason does he have in cracking the book open in the first place?

I am confident at this point that AI-generated books will gain some traction in genre fiction. Someone at this very moment is almost certainly already listening to some AI-generated audiobook of AI slop romances or fantasy novels on Audible. Almost equally certain is that people are publishing AI-generated simulacra of what is ostensibly literary fiction. To what end? I am flabbergasted. For all I have thought about it these past few days, I still have no idea.