Will AI applications ever “replace” writers?
It’s a question to take seriously right now. With media companies like Vice shuttering last week, and mass layoffs at august publications like the Los Angeles Times last month, it feels like a vulnerable moment for humans who write. And the well-known propensity for programs like ChatGPT to write passable — if uninspiring — prose is well documented, but with its jaw-dropping new video model Sora, ChatGPT’s parent company OpenAI has renewed our collective sense that AI companies are on a mission to put people out of work.
But high-profile experiments with replacing human writers have gone badly. AI writers have proven to be error machines that create unreliable junk. So far, even the businesspeople literally going about replacing writers with AI — like Serbia’s Nebojša Vujinović Vujo who told Wired he’s “not a fan of AI” — seem ambivalent about things like morality and quality.
But journalist and author Stephen Marche, has, in essence, replaced himself with AI, making him a great inside source on what these bots can and can’t do. Marche has been experimenting with AI for years, but a recent project of Marche’s is more radical. Via his very own complex method, which involves a whole host of AI-powered applications, he generated nearly all the prose for a mystery novella called, fittingly, Death of an Author.
He may have created the first decent book that anyone can reasonably call “AI-generated,” but it required him to painstakingly force the AI applications to write the book he wanted — sometimes sentence-by-sentence.
Released by Pushkin earlier this last year as an audiobook, Death of an Author, credited to “Aidan Marchine,” tells the story of Gus Dupin, an academic who finds himself embroiled in a whodunnit when the central figure in his scholarship, an author named Peggy Firmin, is shot dead. The whodunnit evolves into a whydunnit, and (mild spoiler) the “why” is revealed to be the turmoil surrounding the replacement of human beings with AI agents.
Futuristic ideas are woven into a deliberately formulaic storyline, with the characters themselves commenting on the repetition of mystery plot conventions, even as they enact them, and the novella includes an AI unraveling an AI-related mystery, all inside an AI-written mystery. And the title itself is a wink at Roland Barthe’s famous essay “The Death of the Author” about the “death” of the artist as the sole arbiter of a work’s meaning. It’s all exactly as meta as it needs to be when you consider what it is you’re reading.
Mashable talked to Stephen Marche to find out exactly to what degree humans can be replaced by the current crop of AI systems, and exactly what he was trying to do by creating (not writing!) this book. This conversation has been edited for clarity and length.
Mashable: Your PR people call this an “AI-crafted” story. It’s common right now to talk about “AI-generated” this and that. But according to your afterword, you really wrestled with these bots every step of the way. So how do you describe what you did to make this?
Stephen Marche: It’s funny. It’s like an impossible position, right? This is a thing that kind of afflicts the conversation around AI generally. I remember when I interviewed — I forget his name, but he was the head of DeepMind at Google when they released PaLM. They described it as capable of understanding. And I was like, “You know it’s not capable of understanding. Why do your PR people say that it’s capable of understanding?” And he was like, “Well, look, what we mean by ‘understanding’ is that if you tell it to write something — like write something in Bengali — it understands you, right? So it’s not a lie, or anything like that. It’s just that the terms that we use to describe these very basic things are challenged by AI.” And in the case of this book, I’m 100 percent the creator. You could not go right now and press a button on a piece of software and have it generate this. It was my task to manipulate the technology to generate this text.
For what this artifact is, “AI-crafted” is like a compromise. I used AI to craft it. I’m the creator, but something else is generating it.
On the other hand, to describe myself as the author…when we use “author,” we tend to mean the person who created the words. And that would also not be true. For what this artifact is, “AI-crafted” is like a compromise.
I used AI to craft it. I’m the creator, but something else is generating it. Any term that we use is going to be imprecise, is gonna be, actually contradictory even. One of the reasons I wanted to do the afterword is because this stuff is too important for us to, you know, do PR with.
Let’s just be very clear about what the process is here — very specific and detailed about how this was made. It would be folly to think that I made a robot, and then the robot wrote a novel.
Did using AI to make this book save you time?
Well, I mean, I wrote a 25,000 word novel in two months. I’m a pretty fast writer, but there were certain aspects of it that were definitely like…you do just go, “Write a dialogue that has this information in it,” and it just generates it. And then you cut and paste it and move it in. I mean, it is smoother.
It cannot do language with double meaning unless you put it in. You can do that, through Sudowrite and so on, but the writing process here is one in which you have to be absolutely clear about what you want.
But I don’t think labor-saving when it comes to writing — or indeed anything creative — is really the point. The point is: What can it do that nothing else can do?
Your story is carefully plotted with twists and turns, but as you say, that’s not the AI. That’s you. What don’t AIs get about plots?
An AI just does what’s come before, right? What it’s really useful for is, if you’re writing a letter of recommendation for a university student, it’ll write that thing perfectly, because that language is so banal. But if you’re trying to actually work out a plot, all it’s going to do is whatever is come before, in some regurgitated form, which actually doesn’t make for very good plots. This technology is derivative. It is fundamentally derivative, and when you use it, you have to use it to that end to get at something derivative. Right now that actually is a pretty interesting artistic reality that I think is going to be really interesting to explore. And I think I took a step forward with this book, in getting towards what the possibility of derivative art — like willfully derivative art — might be.
But you know, why it doesn’t make good plot…I’ve been using this stuff since 2017 and I’ve never seen any of the AIs make halfway decent plots. I think it’s mysterious. It just doesn’t work. If you use it, you’re like, Oh, it can’t do that. That doesn’t mean that it’s useless. It’s just that I would never want to read any of the plots that it comes up with. Like they just don’t feel good. And, you know, the double meaning thing is really key. Right? It cannot do language with double meaning unless you put it in. You can do that, through Sudowrite and so on, but the writing process here is one in which you have to be absolutely clear about what you want.
So if you want a double meaning, you have to actually put that in the instructions. It won’t do it naturally, like you think it will in your brain, because that’s how you write. So that’s another really fascinating aspect of this process. You have to express everything overtly, you have to name all of the literary tricks that you want to do, and then it can sometimes get there.
In the afterword you have this comparison between a writer using AI and a DJ. Can you spell out the connection between what you’re doing, and what a DJ is doing?
The break was invented in 1973. You know, hip-hop became a real presence, like 10 years later, but it was only 10 years later that they really figured out. For almost 50 years, film was like, here’s a clip of a train entering the station. Like no one figured out, like, if you put them together in a sequence, you can tell a story, right?
That’s the stage we’re very much at with this stuff. We are in the it’s-amazing-that-it-can-be-done stage. The actual artistic form that this is going to take will probably will be creative chatbots, which probably will be controlled versions of the limitlessness of these generative language models.
All they’re using it for now is to cannibalize old forms. Take this; write a novel with it. That’s what I’m doing. Right? Those poor people at those science fiction magazines, where they’re being flooded with, like, artificially generated science fiction stories of no value. That’s a hiccup. I don’t think the process has been established at all.
And I mean, I have my own ideas. I’m making my own experiments. I’m, like, you know, in various working groups trying to figure out what to do with this. Ultimately, we’re going to have to see what works here, and go out and test it against reality. And I just couldn’t be more excited. I just think it’s really electric.
For people who want to do this, like, go. The world is there to be figured out. The people working with this stuff in the tech space, they constantly are uncovering incredibly strange things, that they have no idea why they exist, and I think creatively, it’ll be a similar journey, we’re gonna get to effects that we didn’t know were there.
I mean, even in this novel, I have that. I definitely have effects that I did not know existed that emerged from this tech. I know, it’s there. How it’s going to blossom? What it’s going to actually end up being like? We probably won’t know for maybe 20 years.
You can buy Death of an Author for $4.99 on the Puskin website.
