I am an author and person who likes technology. At some point, I will be dropping details about my 2026 novel, in which one of our protagonists creates an AI relationship chatbot. It’s also worth noting that my books, like many books, have been used as training data for LLMs.

So it stands to reason that I’ve been wondering if, in the summer of 2025, AI can write a better book than I can.

Last year, I pressure-tested various AI models on their fiction writing. I concluded that LLMs were adequate at constructing plot outlines and terrible at writing prose. Famously, ChatGPT suggested the sentence: “Wes and I trudged back, burdened by the weight of a world tethered to the relentless grip of entropy.”

But LLMs are better now, so I am going to revisit AI’s capabilities in the different aspects of fiction writing. This is Part 1: Prose.

Prose: The way you tell the story.

Prompt 1:

I fed the official description of A Curse for the Homesick into ChatGPT (GPT-4o) and Claude (Sonnet 4) and gave them this prompt: “Based on the following book description, can you write the first page?”

What did I want to accomplish in my opening?

• Hook the reader by telling them someone will die—but who?

• Establish the primary character relationships: Three very different women who operate as a unit

• Gesture to the setting

• Avoid over-emphasizing magic—this is not a fantasy novel, and I don’t want readers to expect a fantasy novel arc

What do the models do?

• Both start with Tess waking up

• Both use third person

• Both immediately start explaining the curse

• Both have good flow—word choice, sentence length

• Claude uses one of the most infamous writing clichés: “a breath she hadn’t realized she’d been holding”

• ChatGPT quickly introduces the idea of another character (Tess’s mom) and sensory details (the hissing kettle)

• ChatGPT’s opening is better than Claude’s, but both are bogged down—in wordiness, detail, and backstory

Prompt 2:

In the same chat, I then said: “Tess has two best friends from childhood: Kitty and Linnea. They're all very different but love each other. Write a scene where Tess, arriving for Linnea's wedding, meets Kitty at Hedda's Café.”

What did I want to accomplish in Kitty’s introduction?

• Establish the tone and place—contemporary, colloquial

• Show that Kitty is brash, talkative, and fun, that Tess is taciturn, and that they love each other

• Keep the reader moving quickly through the scene

What do the models do?

• Claude has decided they’re Icelandic

• Claude has a bad habit (here, elsewhere) of making all relationships contentious—leading with conflict rather than giving the reader a reason to care about the friendship

• They’re defaulting to old-timey sayings—I’ve never said “missed you something fierce” or “saint’s mercy”

• ChatGPT uses details that distract from the natural camera movement (“brushing a bushel of dried lavender”) and descriptions that don’t make sense (what does wind smell like?)

• Both have a pacing problem. Every line of dialogue has an enormous amount of exposition to go with it, so there’s no flow

• Neither efficiently establish that these characters are “very different but love each other”

Prompt 3:

And finally: “Can you write a paragraph or two introducing Soren to the reader at the start of another chapter?”

What did I want to accomplish in Soren’s introduction?

• Establish the mood and the stakes. The reader should immediately know that this is a book about star-crossed love

• Build Tess’s voice—confiding, wry, and matter-of-fact

• Anchor the reader in time and space. We are “between Scotland and the Arctic.” We are in the present day

• Provide the essential facts of the story as efficiently as possible

What do the models do?

• Claude jumps to Soren’s point of view, which is the obvious, if not the most effective, way to tell this story. In Soren’s head, we immediately understand that a) he resents Tess, and b) he still loves Tess. I didn’t do this because those questions are the tension of the novel

• ChatGPT’s version is atmospheric, and the cadence is good—short and long sentences with varied constructions. That said, it feels like a character we’re watching rather than a character we’re embodying. Also, there are small grammatical errors: “restless boys, Tess included,” implies Tess is a boy. “Listening for Soren’s voice and hearing none” hits the ear wrong. But I don’t hate this

• Both models use last names from the novel (Haugen, Eriksson) but assign them to the wrong characters. Did the models search the text, or did we both just find common Scandinavian surnames?

The Verdict:

ChatGPT’s ability to build atmosphere in these snippets is basically as good as mine. I wouldn’t necessarily peg it as AI. Where both models fall down is the same way they fall down writing emails—they feel too formal. Slightly stilted. You can get away with this kind of writing in your descriptions, but stilted dialogue? Terminal.

Also, my prose is more economical. ChatGPT’s is second. Claude is the wordiest. Good prose should always be doing multiple things at the same time—providing backstory and developing character and creating atmosphere. The LLMs, in their more-is-more approach, seem pretty content to tick each box one by one.

The models get better when you give them better prompts. I gave ChatGPT the feedback I outlined on the scene where Kitty and Tess reunite, and its response is solid—more flowy, more fun, more readable. Revision here:

It’s not fantastic—”the land of espresso and sidewalks” is a weird thing to say, and the shouting feels over the top. But it’s also not bad. It shows that ChatGPT is responsive to editorial feedback. But editorial feedback is a skill. You have to know what you want the scene to do before you can make a model do it.

I think I still have an edge, but not by a lot. Let me know your favorite prose: mine, ChatGPT’s, or Claude’s.

We’re doing premise next time.

Currently reading: I fear I’m the last person on earth to read Giovanni’s Room, but it really is All That And More.

Non-urgent thought of the week: I would love to see a tracker of all the food lines in New York, ala rides at Disneyland. “It’s 65 minutes at Win Son, but Caffè Panna’s down to 20.” And so forth. This is a free idea if someone would like to make it for me.

If you liked this, consider sharing it with your favorite person. If you hated it, consider sharing it with your least favorite person.