I Tried to Write a Novel Using AI — Here’s What Happened

Did you see the story about the fiction author who landed a Big 5 publisher deal, then the internet accused her of using AI to write the novel and the deal was canceled? Jury’s still out on whether she did or not, but if you’ve spent any time on YouTube lately, you’ve seen the thumbnails: “Write a novel in hours using AI.” Bold text. Shocked face. Guaranteed to make a writer feel like they’re either missing out or being lied to.

I decided to try it and find out what’s really possible. Full disclosure: we’re getting in the weeds here.

I ran the experiment using ChatGPT free and Claude Pro to try to outline, voice-match, and draft a novel. Here’s what worked, what didn’t, and what nobody making those videos talks about.

AI Novel Writing Experiment Rules

• I didn’t generate full chapters in ChatGPT because I’m currently on the free plan and didn’t want to pay (I had a paid account for over a year, so I know Pro’s limitations … more on that later).
• I used the technique of creating a “north star” file that had samples of my writing, the tone and vibe I was going for, and a reminder of the main characters and plot points. It was uploaded periodically to remind the LLM when it seemed to go off course.
• I tried two methods of production:
  • 1. Giving each chapter its own chat thread that started with the “north star” document and an outline of the chapter. This kept the memory for each chapter contained within each chat thread. Theoretically, that would free up enough memory within each chat to remain consistent. (Claude recommended this method, BTW.)
  • 2. Generating multiple chapters in the same chat thread. In this version, I started with the north star and chapter outline, then let it generate the full chapter. When that was done, I kept asking for new chapters within the same chat. I got more consistent results with this method, but it only worked for two or three chapters before forgetting things.

Teaching AI My Voice Didn’t Go As Expected

The first thing I did was train the model to write in my voice. I created a project for the book, fed it samples of my writing, described my style and audience, and prompted it to mirror my writing style.

Full context: I was an early LLM adopter. In 2025, I created custom GPTs that could reasonably recreate my voice for short things like blog posts and video scripts. Each one had a different brand voice to match different projects. That changed when ChatGPT 5.0 was released and it suddenly “forgot” everything I taught it … but I digress.

The point is, I’ve done this before. Older models could reasonably recreate my style for outputs as high as 1,000 or 1,500 words.

In this experiment, it sort of worked. With the ChatGPT free version: results weren’t so great. I ran into memory and usage limitations pretty fast. As expected. Claude paid version: it did OK with short sections. When I tried to generate a full chapter, I got mixed results.

The problem comes down to usage limits and memory constraints. You can’t really maintain consistent “training” across multiple chapters of work. Even when I did the trick of re-attaching the training document to refresh its memory, the style was inconsistent. Sometimes it was close, sometimes generic but OK, and sometimes seemed like an all-you-can-eat word salad bar.

The more distinctive your voice, the more work it’ll require to get your fiction to sound like you. If your writing style is what sets your work apart, LLMs aren’t ready to recreate your voice for the length of a full manuscript. It did a decent job at generating mid fiction, though. If you don’t mind sounding generic, it wasn’t terrible.

Outlining Was Useful (With a Catch)

This is where I found the experiment most useful.

When I gave it my core concept, rough character ideas, and a general sense of the story I wanted to tell, it generated a working outline pretty much instantly. That helped. I’m the worst procrastinator when staring at a blank screen. Structuring work takes me forever, and it helped to have something to react to and interact with. I spent an afternoon spitballing and exploring tangential ideas. By the end of that planning session I had a five-book series planned. (Laughing, but true.)

But the AI didn’t do it for me. The outline it produced had plot holes and suggested chapters that didn’t need to exist. There was padding that didn’t add anything. It didn’t consider narrative momentum. (And yes, I prompted it to use the Save the Cat! beat mapping process.)

On the plus side, it was able to help me consider different angles, see subplots I hadn’t thought of, and generally it functioned like a writing partner.

Here’s another thing it did: I asked it to compare my idea to similar titles and walk me through audience expectations and market drivers for the genre. It came back with marketing ideas, tropes, things that work, and whether the idea should be tweaked to better meet market demand.

If you’re interested in selling your fiction, these considerations matter. And it was faster to explore the market this way than digging through comps and sales data yourself.

AI-generated outlines and drafts are useful, but I’d treat it like a head start from a well-read assistant.

Chapter Generation Was the Weakest Link

This is where we need to manage our expectations.

The prose was OK, not great. For the most part, it was grammatically correct and structurally sound. It was also flat. It defaulted to generic dialogue and descriptions. There was a noticeable lack of sensory details, so it didn’t feel “real.” That’s to be expected, really. LLMs are basically fancy autocorrect models, not humans. It couldn’t seem to grasp “show don’t tell,” and there was a lack of emotional depth.

That’s all fixable in your edit. The real problem is it couldn’t remember what happened three chapters ago unless you reminded it. Continuity has to be actively managed. By the time I’d corrected inconsistencies, adjusted the tone, and added sensory details, I’d spent nearly as much time editing as I would have drafting the chapter from scratch.

That doesn’t mean it wasn’t useful. Editing a rough draft is psychologically easier than facing a blank page. But it’s also not the shortcut YouTube thumbnails are selling. You still end up doing work to make it worth reading.

Does AI Work as a Writing Strategy?

Here’s what I couldn’t find while doing this experiment: sales data on AI-produced books.

Nobody making content about AI novel writing is showing whether these books actually sell. There’s no transparency around reader response or whether AI-generated fiction is building audiences. Sure, you could clickbait someone into buying a $2.99 romance novel. But if it’s terrible and a reader doesn’t finish it, have you really helped your future self?

Anecdotally, my 72-year-old mom loves those AI YouTube novellas … you know the ones, they always feature some crazy family drama or Jerry Springer-style plot. She knows they’re fake. She’s just entertained by the wild scenarios they present. It’s something silly to listen to when she’s gardening or sewing. They’re the storytelling equivalent of eating fast food when you don’t feel like cooking.

Those videos get tens of millions of views. It’s wild. I don’t know if the channels are monetized, but if they are, they’re making bank.

But will readers sit through 200 pages of AI content? I’m not sure. Reading is a massive time investment, and readers are getting better at detecting AI. The fiction market — unlike audio or nonfiction — runs entirely on a reader’s emotional connection with the characters. Flat characters and prose aren’t going to build an audience, IMO.

I’m not saying it can’t or doesn’t work. I’m saying I haven’t seen the sales numbers.

Because of that, it’s worth being honest about the time commitment required when writing with AI. The hours you save outlining and drafting will be shifted to editing and fact-checking. And, of course, it’ll still require the same marketing as any book.

Whether that trade-off makes sense depends on your goal. Maybe it makes sense for you.

My Verdict on AI Novel Writing

Here’s where AI was genuinely useful:

• It broke through through blank page syndrome. I’m the kind of writer who’ll sit there staring at a screen forever and procrastinate when I’m stuck. The output gave me a starting point, so I could focus on fixing holes and adding my voice.
• Generating the outline was the starting point, then I was able to explore side tangents, brainstorm, and spitball ideas as if I was working with a writing partner. That part was cool.
• When I ran into questions, like I needed to know how to phrase things in a courtroom scene, the LLM came through with explanations on procedure and resources if I wanted to explore further.

Where it fell short:

• It couldn’t sustain my voice over a full manuscript, even though I went chapter by chapter as tutorials recommend.
• There was no plot coherence unless I explicitly told it what needed to occur in each chapter. Untrained writers might not know what’s missing or how to fix it.
• The prose often sounded generic. There was a lack of tension in dialogue. It often “told” rather than “showed.” Overall, the first drafts looked like prose but often read like a word salad bar.
• Several times it “lost the plot” (literally).

TL;DR: AI isn’t an author. It’s more like a writing partner. It has some good ideas but it’s not great at execution. It’s useful when you already know what you’re trying to write; you’re just trying to speed up the process.

One More Thing

Writing with AI comes down to your prompts and managing memory limitations. There’s no magic formula, but most writers aren’t taught how to direct these tools. Vague inputs produce vague outputs.

If you want to skip some trial and error, I put together a free PDF with specific prompts I found useful in this experiment. It’s built for writers, not tech people. No jargon, just what produced better results.

[Grab the free prompt guide here.]

Have you tried writing fiction with AI? I’d love to know your experience in the comments. What part of writing do you think AI could never replace?