The Rise of the Robot Writers?
Do AI writing programs have anything to offer fiction writers?
Last month, an artist I know called me with a mix of panic and excitement. They’d seen the tests of a new AI art generator called DALL-E (get it? Wall-E meets Dalí?) and said it was a “complete game changer.” That it would, in a few years time, completely change how visual artists create art. It wasn’t actually that the AI art was great. But it was probably good enough—and certainly would be good enough in a few years time—to replace human artists for certain tasks like supplying art for header images of online articles. It could devastate freelance illustrator incomes.
What was more interesting to this artist was the possibilities of using AI art for human artists. To collaborate with the machines. While an AI artist might not produce better art than a human, it can produce art far more quickly. It can easily spit out 100 versions of an idea, allowing an artist to pick the best composition as a template for an original (“original”?) work. Or simply touch one up a bit and call it theirs.
I find this very plausible, and yet for years I’ve heard the same claims about AI programs replacing fiction writers and find those… much less likely. I’ve played around in AI writing generators several times and never found anything they produced remotely interesting. So why does AI seem like such a game changer for visual art but not for writing? Even with AI programs produced by the same people?
First, let’s look at some of the OpenAI’s DALL-E art.
Is this great art? Not in my view. And if you see these images in detail you can clearly see a lot of the AI seams. However, if you were thinking of painting a parody of American Gothic with two dogs holding pepperoni pizza the program is giving you a variety of pretty decent composition ideas. Or if you were a magazine editor who for some reason wanted that image for an article illustration you could simply use any of those. Would they grace the cover of the New Yorker? Probably not. But could they work for your average online magazine with a minimal art budget? Absolutely.
(One thing you can note is that, at least as of now, AI is pretty bad at reading text. The prompt asked for pizzas instead of pitchforks and yet each image has both pizza and a pitchfork or pitchfork-like object.)
Again, this is pretty darn good for the prompt! Although again the text interpretation is somewhat poor. These look like robots made of stained glass rather than stained glass windows of robots. Still, it’s pretty easily to see how one could play around with this for only a few minutes and get a decent article illustration, book cover design, or some such.
Where DALL-E gets really interesting for artists, I think, is that one could plug in one’s own work and the program can spit out tons of variations in your unique style. Here’s some in the vein of one of the great French illustrator Moebius.
Now, as someone who is a great fan of Moebius I find all of these to be impressive for AI but bad for Moebius. His work is simply more elegant and interesting—and of course original—than these images. But. I could easily see how an artist could plug in work of their own and uses images like the above and skip much of the laborious sketching and drafting stages. It could indeed be a game changer.
This is a newsletter about writing. So what about OpenAI’s writing program? The writing generators I’ve seen, including OpenAI’s, can certainly imitate some types of writing. They can pretty easily spit out an undergrad composition essay that could fool a teacher rushing through a stack of papers to grade:
This is generic, uninteresting writing. No one would learn anything from reading papers like that. But then most undergraduate papers are generic, uninteresting writing that say nothing new. (Insert “Since the dawn of time, mankind has always written generic essays or assignments in education facilities” opening hook.) I think the above thread’s warnings about the dangers of AI writing for learning are completely correct. Schools will very soon be filled AI generated essays that can’t be caught with TurnItIn or other plagiarism detection software.
Is there a use, beyond cheating, for these essays? I find the ideas presented later in the thread—basically that AI could be used in classrooms to generate student ideas or examples to critique—to be pointless. This AI program is basically regurgitating generic statements in a way that is functionally plagiarism but not detectable by plagiarism software. I don’t see how that’s a useful guide for students when there are infinite examples of more interesting human-written essays to look at.
Here’s what the program spat out when I gave it a craft newsletter prompt:
Again, this is just robot plagiarism by another name. The program can reproduce existing knowledge, but it doesn’t seem capable of creating new knowledge or insights. AI writing programs have already replaced human writers in certain areas where all that is expected is basic regurgitation of existing knowledge in text, such as writing baseball game summaries from existing box office stats.
What about creative writing? Writing that is trying to be original, insightful, and beautiful? I’m not sure there is much there. AI programs can create fun Surrealist sentences or lines in the same way that a drunk person playing with refrigerator magnet poetry at a party can. It will produce some humorously awkward writing and sometimes a halfway interesting image. But I think it’s a long time before AI writing is going to be filling the pages of literary magazines.
That a computer program can write the above might blow the mind of someone twenty years ago. I’m not sure it’s better than what an average elementary school student would write.
What interests me the most as a fiction writer is whether AI writing programs could function as a “game changer” for authors in the way it might for artists. Could the provide sample structures, plot beats, or character dynamics in the way that they can creative draft compositions for visual art?
I’m not seeing it. And I’m not exactly sure the reason why. Perhaps the programs are constructed differently or maybe there is a larger store of images to work with. But every AI generated work of fiction feels as generic and uninteresting as the student essay examples, whether your prompt is a boring or bizarre one:
For some reason the program crapped out a few sentences into the baboon doctor prompt—and the text it did produce is just recycling the prompt. The first text isn’t in the style of Faulkner at all, but it at least seems to be somewhat in the form of a short story. Still there’s nothing interesting here. I ran this prompt or similar ones a few times and the variations were all generic. For example, the second time the terrible secret the husband tells the wife is that he lost his job. If you’re a writer looking to use AI for ideas, you need it to spit out surprises not the most obvious plot points.
The contrast with the visual images above is, to me, pretty drastic. The AI generated images could with minimal tweaking be published in an online magazine. But the AI short story could not. I think plenty of people would enjoy the robot stained glass images, say, but few would read 20 pages of that generic married couple story.
It seems that while the programs can mash up images or words, they don’t have any real ability to do much else. It’s not going to give you any ideas for structure, pacing, atmosphere, dramatic conflict or really anything at all. At least not anytime soon.
Of course, AI programs are getting better all of the time and at an exponential rate. Still, I wonder if they will ever be useful for creative writers. Let’s say that down the line you could plug 50 Stephen King stories into an AI program and it could spit out a dozen Stephen King style short stories with a prompt like “zombie duck” that were the fiction equivalent of the the Moebius images above. Not amazing, but maybe passable rough drafts for a high school creative writing class.
Is there a use for that, from an author’s point of view? I’m not sure. There are some fundamental differences between visual art and prose art, especially in how they function in the marketplace. A novelist will write one novel (at a time) and sell it to as many people as possible. A gallery artist can only sell one painting to one person. There is an output pressure on visual artists that doesn’t exist for novelists. A painter might easily paint 100 versions of a general idea to sell to 100 people, while a novelist is trying to create the single best version of one idea at a time.
Authors have far more ideas than they have time to complete. If story ideas or plot outlines are the equivalent of different image compositions, authors don’t really need help with those. Most of us can come up with the outline of a story in a few minutes. The issue is the writing of the ideas. Of fleshing them out into complete books.
On the surface this seems like exactly what AI would be good for. Currently this is what AI writing programs are the worst at. AI programs don’t know how to build upon existing work, allowing meaning to accrue over the length of a work. They can’t do plot or character arcs. The heart of fiction writing lies in this accrual of meaning. Of getting into the text itself and finding connections and building meaning out of the progression of language. AI programs spiral into repetitive gibberish the longer they go on. But let’s say they fix this issue in the future. A fundamental issue remains. It’s not writing. It’s reading.
First, I don’t think people will ever be interested in robot writing outside of the novelty. We go to writing for things like learning about the experiences and perceptions of people who are not like us. To see in new ways through the eyes of other people. But more than that, there isn’t really a need for an increased book output.
There are already far more novels published than anyone can read. Far more manuscripts filling agent inboxes than will ever be published. Even a novelist as famously productive as Stephen King only averages a book a year or so. Could an AI program help King produce 100 novels a year? Maybe. But it’s unlikely people would read that many, and even less likely that the publishing ecosystem could handle them. How many book reviews can be published? How much shelf space exists in bookstores? How many books can a publicist work on at a time? There are already far too many books written than we can handle. What the world needs is better books. More original books. More visionary books. But not simply more books.
There are some things that AI could do for writers. For example, I’d love to have a program that could switch a passage from third person POV to first person POV or present tense to past tense. That could be a godsend, though it is less about replacing creativity and more about speeding up the process. Closer to “find and replacing” a character’s name in Word than using someone else’s plot outline.
So for now, as a teacher I’m pretty worried about the havoc AI essays are going to cause. But as a novelist, I don’t think the robots are going to replace or even help us anytime soon.
As always, If you like this newsletter, please consider subscribing or checking out my recently released science fiction novel The Body Scout, which The New York Times called “Timeless and original…a wild ride, sad and funny, surreal and intelligent” and Boing Boing declared “a modern cyberpunk masterpiece.”
Feels more useful for drafting SEO-driven "content" for businesses than anything.
Enjoyed this piece! I agree with your sentiment about the AI writing tools. I’ve played around with some of the AI tools like Sudowrite (a tool built on top of GPT-3), which are designed to assist rather than replace. I wrote about my experience here, last July: https://letter.salman.io/p/boredom-and-creativity?s=r
Broadly speaking, I too have not found a place for them in my writing process. But I do believe they can help folks bust out of writing blocks by seeing directions they haven’t thought of. I see their greatest potential as idea generators rather than word generators.