Author Stephen Marche on how to create with AI without outsourcing the fun part

When you last collaborated with a human, you probably had to work around some obstacles: Finding time on your calendars, agreeing on goals, negotiating conflicting points of view to get to common ground. But when you’re collaborating with a machine, the process can get a little more… mysterious.

Ask Stephen Marche. He’s been immersed in the highs and lows of co-writing with people and machines for years. This month, Akashic Books is publishing his new book, The Last Election, a political thriller co-written with former presidential candidate Andrew Yang. Earlier this year, Pushkin Industries released the audiobook Death of an Author, a Marche + AI mystery about a literary scholar entangled in a murder investigation involving a novelist who was the focus of his studies. Imagine tech noir à la Blade Runner as told by a chatbot trained on Raymond Chandler.

In his essay in The Atlantic, Marche describes how he produced the story of Death of an Author by experimenting with different prompting techniques using a combination of ChatGPT, Cohere, and Sudowrite. He began with instructions to write in the style of various authors and publications. Then he’d switch tools to rephrase the original output, and keep editing until each passage sounded the way he wanted.

Marche was an early adopter of AI as a writing assistant, but he wasn’t immediately sold on its potential. In fact, his first piece on the topic was the 2012 essay Against Digital Humanities that critiqued Google’s efforts to convert literature into data.

So how did he go from skeptic to advocate? We spoke with Marche to hear about that 11-year journey and find out how he became optimistic about the future of human/AI collaboration. 

When did you become convinced of generative AI’s potential? 
I wrote my first algorithmically generated story for WIRED in 2017. So I’ve been doing this, in terms of AI writing, prehistorically. But I got a computer scientist to build me a machine to do that. Then I got early access to various Cohere products. I've been very proactive about this subject. What began as resistance has become, “Let's see what can be done that’s really valuable.”

In the Afterword of Death of an Author, you described your technical methods. But how did it feel—in terms of the creative process—to prompt a story compared to writing one? 
There were sort of two separate things. One was, I was training the device to produce what I wanted it to produce. That was with particular stylistic metrics using these three different techs. That was me learning how to use the tech, but also testing the limits of the tech, and getting right to the edge of what it could do—because it was early days for ChatGPT. That was some of the best AI that's ever existed. So feeling and seeing what it could do was really impressive. But there were also times where the machine was creating something that I would never have created. And those are the really interesting moments, when the machine was generating something—I don't mean to sound weird about it—but it was like I was getting an alien to write a short story.

Can you give an example of something it wrote that was alien to your way of thinking? 
At one point, I got it say, “She walked down the hallway like an LP going back into a sleeve,” which is a great line. But it's more like, I'd be doing a scene of a woman going on a dock somewhere and it would give me a description like, “She's wearing a yellow dress.” It was very specific, like it had an idea. Like, “Let's make this character this way.” That's an illusion, obviously. There were only 20 or 30 moments like that in a whole book, but when they happened, I really felt them. 

When you collaborate with generative AI, how do you keep from outsourcing the most enjoyable part of the job—the creative discovery?
The nightmare world is the world where machines write the novels and the people drive the cabs, right? I would say, from my experience in using this, that's the opposite of what's going to happen. This is a creative tool like Photoshop or CGI. In linguistic terms, what it’s good at is producing the standard form of something. So if you want a recommendation letter, there's nothing better than ChatGPT. But it is not an original mind. Its intelligence is of a reproductive variety. 

This may sound optimistic, but I really believe the actual originality of thinking is going to be increased in value. Much like what happened with painting where, when the photograph comes around, the ability to reproduce what a face a human face looks like suddenly loses its value because the machine could do it. But the ability to make an interesting painting doesn't decrease because of that. In fact, now you need to be more wild with it. I think that's something that could absolutely happen with this technology. That's my hope. I live in hope of that.

You’ve said you expect personal expression to become more valued in the future. Did the process of creating a novella via AI make you feel liberated to explore topics you might not have explored if you were authoring under your own name?
For the purpose [of creating Death an Author], we were doing an experiment. I've done multiple experiments in the space. And they are experiments—I don't think anyone's going to do this as an actual creative process. It's more like, let's see what can be done. 

But finding the exact form, that's going to be the trick [for AI creativity]. In film, that arguably took almost 30 years from the invention of films to the development of a real narrative structure that people could watch. I think it's gonna take us a long time to figure this out. But it is, for sure, a new medium. And it's exciting to be at the beginning of something rather than the end. 

But you know, I write novels and I love the novel, but the novel is in an archaic form. It's been around for a long time. And the best examples of it have probably been written. But the best examples of whatever AI becomes—and God knows what it will become—have yet to be written. So that's the source of excitement.

You’ve previously compared AI to the invention of hip-hop. Do you see AI-generated art as something like a new genre that will exist alongside human-generated art?
Think about Google’s life coach. What that is, and we don't want to admit it, but that's an art form. That’s an artistic expression. That is going to be infused with the values of its creators. They [also] don't understand that there is no such thing as “the life coach.” There have to be multiple life coaches. So what they're making there is a new art form. It's a new illusion with language. One thing that’s fascinating about AI at the moment is that people don't understand that it is just an art form. It's a bunch of tricks with language that amounts to a new form of expression. The people who understand that are going to be the ones who actually make the meaningful stuff.

You’ve said chatbots give you a different way to think about a problem. How does getting input from an non-human perspective compare to getting feedback from human editors?
I think it's roughly the same thing… If I'm going on a trip and I ask it, “What should I do there?” It's incredibly good at that. The next thing you do is you check that against the person. You don’t believe it in itself—you have to go and check it out. [For] the people that I trust that I interview about this, human/AI collaboration is the name of the game. That's where the really fruitful material is gonna come from this. 

Would you be curious to see an AI-generated film version of Death of an Author? 
I would be totally into it, yeah. First of all, I'm totally on the side of the [WGA] strikers, but I don't really think they've understood how AI isn't going to replace anyone—certainly no more than CGI did. It's a very cool potential tool. Any studio executive who thinks they're gonna hire AI rather than a writer to create a script is out of his mind. That's just not what is going to happen here. But because of that, there's this fear of AI. People don't necessarily want to do it because it's got a kind of nasty representation [as] a replacement for human creativity. Whereas I really do think of it as just an extension of human creativity. I don't think it's going to replace anyone.

There seems to be a growing appetite for authenticity in the wake of deep fakes, filters, and auto tune. What's your feeling about this? Do you have some sort of process for maintaining a sense of authenticity when you're creating with AI tools?
The algorithms that drive our culture are so profound and they're invisible. That's their power. They’re hidden from everyone. When you [use] AI, you're consciously grabbing hold of the algorithm as a tool of culture. And that's its real power. I'm not a very optimistic person, but I actually do have hope for this. I think there's going to be a use of this as a way around the algorithms because nobody knows what it is yet. No one knows what the next day is going to be. So that's exciting.

What are you most excited about doing with AI next?
I've been working with the guys at Cohere to come up with an infinite story. It’s the same story, but because it's trained on prompts, it generates a different version of the story literally every time you open it. It has a perfect hermeneutical system where, if you asked it to write something in the style of the 18th century, it doesn't ever use a word that came later. Why? No one knows! That alone is an incredibly fascinating question.

I want to try and use it to fill in the gaps in some unfinished or broken manuscripts from literary history, particularly The Epic of Gilgamesh. I've been trying to use it to recreate a version of the Oracle of Delphi. There's so much to do with [AI] that’s so interesting. For anyone who cares about language, it is absolutely the most interesting question of the world—how to use it, what it is, what it means. 

This interview has been edited and condensed.