This started as a voice note, got cut off when Chris called, and then got a second part after I gave all three birds cookie time, because they will not let me finish a thought until everyone has been a good girl or boy and gotten a treat.
So this is a thinking-out-loud post. I want you to argue with me.
The lunch that started it
I have a standing Zoom lunch with my friends in St. Louis. I mentioned I’m writing a book, and one of them pointed out that AI-written books can’t be copyrighted. I already knew that. It’s part of why I’m not shopping Hold Without Taking to a big publisher.
The actual rule is more nuanced than most people think. There’s no magic percentage. The U.S. Copyright Office says fully AI-generated material can’t be copyrighted, and in a mixed work only the human-written parts are protected. You have to disclose the AI parts when you register. The Supreme Court declined to take up the challenge earlier this year, so for now that’s settled.
Hold Without Taking is about 62% AI-written. The idea, the world, the backstories, the parts I wrote and the choices about what stays are mine. The other 62% belongs, in a sense, to nobody.
That got me thinking about where book writing is going.
My hypothesis: books are going the way of open-source software
There’s a shit ton of software out there that’s open-source code with a thin layer of modification and a huge marketing budget. Lifetime deals, Facebook ads, the whole thing. Some of those companies make real money. A lot of them probably don’t. But the product isn’t really the code. The product is the marketing.
I think AI-assisted books are heading the same direction. If a big chunk of a book can’t be owned, the value moves to whoever can put money up front and get it in front of readers. I think we’ll see agencies act as the marketing machine and take the financial risk, the way publishers always have, except the “manuscript” part gets cheap.
AI book writing is basically adding to the public domain, just earlier
Here’s the comparison that really clicked for me. Penguin Random House makes money republishing public domain books. I could download Little Women from archive.org for free right now, or some obscure family genealogy nobody’s read in 80 years. (I’m very slowly digitizing my great-uncle’s family history book, Our Seven Children, over on VintageReveries, so this is on my mind.)
Penguin still sells Little Women because they’re better at production and distribution and reaching people than I am. With ebooks, production costs are almost nothing, so it really comes down to marketing.
If AI-generated text can’t be copyrighted, we’re essentially pouring new material straight into something that behaves like the public domain. The difference is that public domain normally happens after an author has been dead a long time and the copyright runs out. These books were never owned in the first place.
So what happens when someone takes the AI-generated parts of your book, puts their name on it, and outmarkets you? Does the loudest marketer in the room win? What does ownership even mean at that point? I don’t have a clean answer, and I don’t think anyone does yet.
Why I’m running these experiments
This is the question underneath everything I’ve been testing. If you’re one of the people buying book-writing software because the Facebook ads look cool (hi, I’m one of you), can you actually produce something worth reading? Is the tech mature enough?
Here’s the test. I’m having a lineup of models each generate The Affirmation Glitch as a fiction book, and I’m judging which one does it best. When I publish the results, I’ll tell you which model won and why I think it was the best, which ones came in close behind, and which ones were the worst.
I’m running everything on BookyAI’s basic settings on purpose. BookyAI has shipped tone and consistency improvements that I haven’t even gotten around to evaluating yet, and I’m deliberately not using them for these tests. I want this to test the model, not the tech wrapped around it.
Quality isn’t the only thing I’m measuring, though. There are real trade-offs on time and machine. Will the winning model need my 4070, or will it be one I can load on my 2060 if I’m willing to let it render over a long weekend? And on money: will the winner cost less than $1 total to generate on OpenRouter, or will it look more like the roughly $50 in Fable API costs that Hold Without Taking ran me? That was on an older version of BookyAI, and honestly, even with the basic features, it felt like being in a house with no doors.
BookyAI keeps shipping features that almost make me want to scrap my quality check. But people have been emailing me about my BookyAI posts and reviews, and it’s clear there’s a real need for a plain-language quality comparison of local and low-cost models. So I’m finishing it. Preliminary results are close – I keep pulling myself back from overcomplicating things and getting down very interesting rabbitholes.
You can probably tell I’m still a big BookyAI fan. Or maybe I’m just stubbornly committed to it, since it’s the first AI book-writing software I’ve tried. I just got access to some others, so we’ll see. But honestly, whoever makes it, the tech is close. There are probably 20 companies you could find easily doing this, and hundreds of people vibe-coding their own versions with Claude Code on nights and weekends. That alone excites me.
Will anyone care that AI wrote it?
Here’s the question I really want to ask: if the writing is good and the premise is interesting, will anyone care that it was written completely by AI?
My bet is no, not in the long run. AI is getting so good at prose (if it isn’t already there) that people will start assuming everything is written by AI. And once everyone assumes that, “was this written by AI?” stops being a useful question. The things that set a book apart become the idea, the originality, the editing and the marketing, not the skill it takes to sit down and produce 127,000 original words.
I know some readers care a lot right now, and I get why. But I think that’s going to shift faster than most people expect. Tell me if you think I’m wrong about this one especially.
I have a queue of other book ideas competing for my attention right now. Which is exactly why I need to stick with this quality check instead of chasing the next shiny idea. If I’m going to hand my ideas to a model, I want to know which one actually deserves them.
But I don’t think anyone else would’ve come up with the world behind Hold Without Taking. That idea is 100% mine. And without AI writing 62% of it, it wouldn’t exist. It would’ve taken me months to hit 127,000 words, and I get distracted, and life happens.
That’s the part I think gets lost. AI is giving people with good ideas but not a lot of time or resources a real leg up. It’s leveling the playing field in a way.
The slop problem is real, though
There’s going to be a lot of garbage. AI recycling other people’s ideas, models trained on AI output, and plenty of people with small imaginations and big marketing skills putting out bad books and making them the public face of “AI books.” We’ve already watched this happen with make-money-online advice on social media. AI will just speed it up.
It’ll get harder to filter, and people will probably use AI to help filter it. And that’s where the best marketer ends up competing with the best book.
A quick note on local models and the environment
People running models like Llama on their own machines are using very little energy compared to the big cloud data centers. If your house (or even just your computer) ran on solar, your footprint would be close to zero, aside from the energy it took to train the model to begin with. We’re not solar, but it’s worth saying in a conversation that often treats all AI use as the same.
Two things can be true
None of this is purely good or purely bad. Two things can be true at once, and sometimes five things can, depending on where you stand. I think we need to be careful about getting moralistic about it. It’s changing things, and I’d rather understand the change than yell at it.
Full disclosure: how this post got written
This post was generated by Claude Opus 5.5 from voice notes I recorded in my garden. The style instructions came from Gemini Pro in Google Workspace, which I had pull my voice out of 12 years of emails to a former pen pal. So AI did the drafting, but the writing style is mine. (Which also means, by everything I just said, parts of this post probably can’t be copyrighted. Fitting.)
I’d challenge you to try it: train AI on your own most candid personal writing (not necessarily your “best”), maybe on your diary, and see what comes back.
Tell me I’m wrong
This really is an invitation. Tell me I’m wrong. Tell me what I’m missing. Send me the thing I didn’t know. I might have comments turned off, so contact me thru my contact form here instead.
I think this is an exciting time to be alive and to be experimenting. This is the why behind everything I’ve been doing.
0 Comments