We made thousands selling AI-generated books
Can artificial intelligence write entire books? What we found.
lostbooks.medium.com
Many creative people have been seeing the writing on the wall: AI is coming for your jobs. That is, if you’re one of the lucky ones able to live off your creative work in the first place — which precious few are. We actually think, however, that this is a needlessly fearful and misleading trope, and that the reality is AI content generation tools can be a tremendous boon to writers and creators of all stripes.
In fact, we’ve made several thousand dollars by incorporating these tools into our workflows after pivoting Lost Books into AI publishing in August of this year. Our sales went from maybe 1–2 units in a month if we were lucky (all too common for indie publishers), to 1–2 units per day, on up to 10+ units per day on good days, where viral Medium Partner posts drive ebook sales and earn money on their own.
So how do we do it?
You can peruse our AI ebooks here to get a feel for what kinds of products we are offering:As of this writing, we have published 32 titles since the beginning of August, and have one new finished volume we have not yet put up for sale. Which works out to a new book around every 3.5 days, or two per week.
The genre we publish in might be described as something like AI surrealist trash pulp sci-fi conspiracy world-building. It is a hyper-specific niche, which seems to have a great deal of reader interest. The books consist primarily of exposition, world-building, and in-universe lore. They are heavy on details, and light on narrative, plot or characterization. In other words, everything they tell you to not do as a writer — and yet they’re selling like hotcakes!
Depending on the length, complexity, and topic, each book takes between 6–10 hours to produce, though a few were even faster. In general, the lengths range from 2,000 to 5,000 words, with a handful of experiments outside that range. Each book also includes usually somewhere between 30–150 AI generated images. They sell for $1.99 or 2.99 USD each. We only sell on Gumroad, and nowhere else, and only do promotional posts about them on Medium, with very occasional links posted to Reddit. And that’s it. The rest of the money came from Medium Partner posts.
Now, some readers might be thinking: two thousand words? That’s not even a book! That’s ‘just’ a short story. If you’re in that camp, do you mind if we ask how much you’ve made selling your short stories, and how many units you’ve moved in a given time period? When people buy one of your pieces, do they often come back later on and buy 6 or 8 or 10 more? Because that happens with our “not books” with a high degree of regularity.
We’ve sold ~250 books since August, and given away another fifty in free downloads to bring new users into our ecosystem. You see, all our books cross-reference each other: so when you’re rabbit-holing into our “not narrative,” you end up stumbling onto another and another and another book that expands outward from whatever your starting place is as a reader. If it’s the kind of thing that piques your interest, you’re bound to buy more, as so many of our readers are proving.
Transmedia storytelling & Networked narratives
As Milorad Pavić wrote regarding his exceptional tome, Dictionary of the Khazars:Such is also the case with our AI lore books. Though there might be a handful of ‘keystone’ titles, there is no “official” entry point into the universe — or more properly multiverse — we’re building. And no path every reader must follow between them. Henry Jenkins way back in 2007 was calling this transmedia storytelling:
“Most often, transmedia stories are based not on individual characters or specific plots but rather complex fictional worlds which can sustain multiple interrelated characters and their stories. This process of world-building encourages an encyclopedic impulse in both readers and writers. We are drawn to master what can be known about a world which always expands beyond our grasp…
Transmedia storytelling expands what can be known about a particular fictional world while dispersing that information, insuring that no one consumer knows everything and insur[ing] that they must talk about the series with others… Consumers become hunters and gatherers moving back across the various narratives trying to stitch together a coherent picture from the dispersed information.”
A related concept in narratology is the networked narrative, via Wikipedia:
Unity & Coherence
All of that is to say that AI text and image generation tools are incredibly well adapted to this kind of hyperrealist storytelling which, to quote Wikipedia’s networked narrative page further, are “defined by their rejection of narrative unity.”Which, to phrase it differently is: AI writing tools are currently terrible at tracking narrative details, plot elements, character arcs, etc. They just suck at it. But they are going to get much better quickly. This you can bank on. (That said, Subtxt is an AI writing app which claims to be good at that today; we have not tried it as it seems to require writing within a very specific framework that our experimental works just don’t fit within).
So can they write a whole book? Yes. Can they write a coherent book? It depends on what your definition of “coherent” is, and what your definition of “book” is… For us, the answer is “yes, kind of.” Though we will admit that for many writers more married to convention than ourselves, the answer may be a more ponderous and frustrating “not really/not yet.”
For our part though, we’ve taken inspiration from Argentinian novelist Cesar Aira, on whom Wikipedia states:
“Aira has often spoken in interviews of elaborating an avant-garde aesthetic in which, rather than editing what he has written, he engages in a “flight forward” (fuga hacia adelante) to improvise a way out of the corners he writes himself into.”
For the work we’re publishing, we have intentionally embraced what many might consider be the “flaws” of AI writing tools: their poor coherence, their inability to track narrative arcs, their tendency to just invent “facts,” and so on. We like it. We want to recognize these artifacts from our tours in the Uncanny Valley as a feature not a bug, and as something that points us towards new multiverses of storytelling we’re only beginning to scratch the surface of. We think of this as something like the Japanese idea of wabi-sabi, where the imperfection is part of what creates the magic and the beauty of the finished works. The role of the artist here, we think, is to effectively interrogate these technologies, and use them to reflect their own true natures, and ours.
What does that mean in a practical sense? It means that from page to page or chapter to chapter, and definitely from book to book, the details of the story presented may not agree with one another or totally add up. They may outright contradict, or they may offer an alternate parallel universe explanation of something that seemed already settled. The effect on the reader is disorienting, challenging, and provocative — and yet, some people seem to like it so much they come back to buy again and again. So how can this be wrong? (Of course, this is the internet, so no doubt dozens of people are lining up with their cudgels to expound on exactly why this is wrong…)
Tools & Attribution
We haven’t spoken yet of the specific tools we use. The majority of the books we’ve published using AI have leaned on the writing models available through the TextSynth website for free. The models we favor are all open source there: GPT-J, Fairseq (made by Facebook), and GPT Neo-X.We tried GPT-3 (OpenAI) and found its generations to be a bore for our purposes. Too vanilla and not creative enough. We experimented in Sudowrite, and think it has some potential, but generally didn’t like the overall UX of it — though it’s quite popular. We dabbled in a variety of others that are more explicitly geared towards marketing and SEO copy, and found we categorically hate them. Many of them rely on GPT-3 as the underlying model, first of all, but second of all are designed to just churn out commercial drivel wholly unsuited to our weird ends.
We’ve also been making heavy use of Character.ai to both generate dialogue-based pieces, but also bots we set up specifically as writing partners to flesh out ideas. It’s hands down one of the best tools out there, and one of the best uses of that tool.
For images, we rely on OpenAI’s Dall-E 2, and were for a time using Stability.ai’s DreamStudio to interface with the underlying Stable Diffusion (SD) model. We don’t especially care for the company behind Stability.ai though, and found you can get tons of free SD images via a third party site called PlaygroundAI, so we use that instead now since it’s free and their UX is much better.
Our workflow for a new title usually goes something like:
- Come up with a title that fits in one of our several title schemes
- Flesh out 4–5 bullet points of what the concept is
- Start generating images around that concept and see where they lead
- Write a few sentences into TextSynth, see what different models will generate from that, and then copy paste any good results back into the text box, and generate some more
- Iterate on all the above
- Edit the results to keep things on track
- After generating some more finished text, use evocative quotes from that to generate more images for the book
- Dump the generated images into Adobe Lightroom Classic to organize them (an absolute must for managing hundreds/thousands of images)
- Dump the texts into Vellum, an excellent ebook writing app
- Refine which images will be used in the final, and then drop those into Vellum
- Go back through our text, and link back out to other books with more info about key points in our multiverse
- Add an “Also by…” list at the end of each volume with all the other books in the series
- Pull a generated image & upsize it to use for the book cover image & add text in photoshop
- Create a preview grid of sample images contained in the book for use on Gumroad
- Generate epub & mobi files out of Vellum
- Upload all the files as a new product on Gumroad & write a very short sales pitch for each one that fits our niche
- Publish a promotional Medium Partner post about the new book & save it also onto Reddit (PS. these promo posts work best when they themselves can be viral posts — not always easy or predictable, but can earn you money on top of unit sales)
- Collect $$$
We don’t specifically list all the AI tools we use in our books (because most readers probably don’t care or know the difference). We do, however, flag within the text of each book that artificial intelligence tools may have been used to edit, enhance, or generate aspects of the content, which is (we think) in keeping with OpenAI’s publishing guidelines and the recommendations laid out by the Alliance of Independent Authors, regarding ethical use of AI tools — even though we are not a member of that organization.
So, will these tools “steal the jobs” of writers and artists?
We opened with that question — or more properly that fear — that creative people will be driven out of business by AIs. The feeling of fear might be real, but we believe it to be unfounded. What these tools do instead is radically alter the possibilities available to creative people to plug into their workflows.That is, they change the locus and scale of creative activity. Instead of creating one picture (which you can absolutely still do the ‘old-fashioned way if that’s what you love; no one is stopping you), you get to create one thousand. Instead of writing 500 words at a clip, you might be able to pull off five thousand. If that doesn’t interest you, then stop right here.
But if any of that does interest you, then by all means, go out and explore. Get in on the ground-floor, and become one of the early pioneers who goes on to do great things with the technology as it evolves — and in so doing, actively helps it evolve in directions that will serve creators and audiences.
You might even earn some money along the way if you’re lucky, like we did.
