When Will Generative AI Write Better Books Than Humans?
And it is when, not if, in most scenarios.
OpenAI’s ChatGPT exploded across the internet in November 2022 and is doing for text what DALL-E and similar generative AI models began doing for images in 2021. Creatives in most fields are rightfully taking interest. Many are rightfully concerned. But what about authors?
A great pathology of modern times is the mistaken belief that everyone’s opinion matters about everything. So, for what it’s worth—I’ve been a technologist for the past 15 years, published two science fiction novels, and read hundreds of books across many genres. Here are my current thoughts. I’ll throw in some jokes along the way.
Let’s go!
And let’s further expand the title question: When will generative AI write which books better than which humans, according to which other humans? And let’s start with the last piece because it influences the others.
Input -> Output -> Input
Generative AI models are trained by ingesting an enormous amount of data and then producing “new” content—more on this later—that is scored by humans so it can improve (more fascinating details here). It took lots of time and money to train publicly available models like ChatGPT before they were released into the wild, and it still takes time and money to produce useful output.
The scoring for some types of content like math and historical facts is fairly objective—as much as facts are still objective—while for others like images it’s more subjective. Books are subjective, but it’s helpful to distinguish between the quality of execution and whether the actual content—ideas, plot, characters, etc.—matches your taste as a reader. The former is (subjectively) less subjective; there are many books I recognize as well-written by competent authors but I have no interest in reading because they don’t align with my personal interests (ahem, Colleen Hoover). The real bar for generative AI is producing books above some acceptable standard of quality and execution with content that is of interest to enough people to make it worth the incremental creation cost—something humans can indeed score. Thus the subjectivity barrier for books will be overcome just like it has been for images.
(There is of course a post-modern perspective where nothing is better or worse than anything else and everything is about power dynamics and whatnot. But that perspective can be safely ignored by those trying to do useful and interesting things during their flash of sentience between twin infinite voids.)
The bigger challenge than subjectivity is training feedback loop length. Someone can score an image in seconds and a paragraph of text in less than a minute. But it might take 5–10 hours or more to read or listen to a book, and much longer than that to fact-check it. And most books require a complete reading before their content can be fully evaluated; what if you get 90% of the way through a thriller or mystery novel only to find out the big twist makes no sense or is missing? What if dozens of important loose ends never get tied up, or there are contradictory facts and conclusions?
Thus the time and cost to properly train a model that produces quality books will be significantly greater than today’s models. That doesn’t mean it won’t happen—Google brute-forced Street View by slowly driving their cars down essentially every road on the planet—it just means it will likely take years, possibly many. And like today’s generated content, the AI will likely produce the first draft—the hardest part—and the author will then serve as an editor and fact checker. (this will also be true of AI-generated movies)
Dewey Decimal Decimation
Some kinds of books will be easier than others to get over the quality bar using generative AI. Let’s start with the highest-level categorization—fiction vs. non-fiction.
Fiction
Before breaking down fiction further, let’s return to the “new” content generated by AI. These models only create iterations of what they’ve been trained on. The output may combine various elements from the training data to create something technically “new,” but it won’t write a story unlike anything written before.
But does that matter? Almost nothing is entirely original—most novels employ common plot structures and character archetypes and are derivative of others. Furthermore, there’s already an effectively unlimited number of books, of which anyone has only read a tiny fraction; the generated output only needs to be new enough to the reader, not truly distinct from the entire corpus of fiction.
Fiction further breaks down into categories based on:
- Length. Per the training feedback loop discussed earlier, it’s reasonable to expect generative AI to proceed in this order: flash fiction -> short stories -> novellas -> novels -> series.
- Audience. Given the content’s relative complexity, it’s reasonable to expect generative AI to proceed in this order: children’s -> middle grade -> young adult -> adult. I have a 20-month-old son and am something of a children’s book connoisseur right now. These books contain so little text that I bet ChatGPT combined with one of today’s generative image models could produce good enough content. (perhaps I’ll give it a try…)
- Genre. This category is trickier. Some genres—like romance, the best-seller—can be very formulaic, and others—like fantasy—include mostly made-up things. I think these will be the first genres AI learns to create, along with overly avant-garde literary fiction, which ChatGPT could probably produce today. My favorite—science fiction—and the genre I write in—near-future technothrillers—will be more challenging. And the “harder” the science fiction, the harder it will be for AI to generate because of the required scientific and technical accuracy, which ChatGPT famously lacks (though here is an interesting piece on incorporating WolframAlpha to solve this problem). Per the above discussion, genres like mysteries and thrillers that rely on twists to tie everything together will also take longer.
Beyond these fiction categories, writers broadly fall into two groups:
- Pantsers: Those who write by the seat of their pants and just let the story flow. Stephen King and many of his infamously unsatisfying endings are a canonical example. It will be easier for generative AI to write these kinds of books.
- Planners: Those who intricately plan out their work ahead of time. George R. R. Martin is a canonical example. These novels will be harder for AI.
Nonfiction
Non-fiction also breaks down into many categories, but one overarching problem that doesn’t exist with most fiction is the need for some level of objective accuracy. You simply cannot blindly trust the factual content of today’s ChatGPT. This will of course improve over time, perhaps via something like WolframAlpha as mentioned earlier, but right now it would be faster to create a textbook from scratch than create one via ChatGPT and fact-check the whole thing.
But what if you trained a generative AI model exclusively on textbooks? This could be one way to resolve factual inaccuracy, but it will also run into the same issue as fiction: These models aren’t really creating new content; they are entirely reliant on the content on which they’ve been trained.
Generative AI may be able to accurately summarize existing information. But it’s not clear how anything like today’s models will be able to put forth the kinds of new and interesting hypotheses that led to game-changing books like The Black Swan, The Selfish Gene, or Guns, Germs, and Steel.
And what about some of the most popular non-fiction categories—memoirs and (auto)biographies? Many of these “authors” employ ghostwriters, which is a promising use-case for generative AI, but it obviously can’t write about never-before-documented aspects of someone’s life and thoughts. Once again we’ll likely end up with a human-AI team outperforming what either could do solo.
Now, plenty of non-fiction books are completely devoid of novel ideas or information. For example, today’s generative AI can likely already produce more interesting works than many self-help and business books. I suspect some already have non-human authors.
Further Considerations and Conclusions
Other considerations will impact timing, like copyright and cost. These are messy but they will get worked out. The more interesting questions are will people want to read these books once the novelty wears off, and how much does human authorship matter?
People evolved to be deeply interested in other people. Just look at the bizarre phenomenon of celebrity. Every generation’s pop music is basically the same as the last’s, but the new generation loves it because it comes from a new crop of artists (although there are already some computer-generated musicians). Few people are interested in watching machines play sports, even though they can vastly outperform humans.
Excepting some non-fiction genres where the author’s credibility matters, the actual person behind the art is arguably less important in writing than in other fields. Many authors—like myself—already use a pen name to obfuscate their identity. Maybe instead of reading all your favorite author’s books, you’ll read everything written by your favorite AI model in the future. However, I think it’s more likely you’ll do both.
Humans are storytelling creatures. Across all categories, humans will write the best and most groundbreaking books for a long time. Perhaps indefinitely. Like everything these days, book sales follow a power-law distribution, so it may also be the case that most sales are by humans. But there is a long tail.
Thus, there will always be a place for human-created art, but as the low-cost generative AI version becomes good enough for some number of people, that place will get smaller.
So, the answer to the question of when will generative AI write better books than humans is — like the answer to many interesting questions — a resounding “it depends.” But it will happen, it may just take a while, and it will take longer for some kinds of books than others.
In my first novel, Mind Painter, the eponymous technology allows augmented humans to “create whatever I can dream in real-time, unconstrained by the need to master the accompanying skills. To make music without playing an instrument, tell a story without writing, or create a visual scene without mastering a paintbrush or program.” We aren’t there yet, but I wrote that in 2018, and that part of the book takes place centuries in the future.
If you find yourself cherry-picking examples where generative AI gets it hilariously wrong to make yourself feel better about the accelerating rate of technological change, recognize your confirmation bias. Don’t fight the future. The only option is to embrace and find peace amidst the chaos brought by new technologies.
For example — what if I told you ChatGPT wrote this entire piece? I told you there would be jokes!
