Why I used AI to write my novel

The machines are coming for our jobs. It’s a message we’ve been hearing for years: robots will replace factory workers, self-driving cars will supplant truckers and taxi drivers, and now artificial intelligence tools like ChatGPT will take over everything from teaching to lawyering to penning the next episode of Law & Order. It’s wise to be skeptical of these claims, to acknowledge the cast of powerful corporations that benefit from AI hype. But the truth is that the machines are coming—in fact, they’re already here. Talk to a teacher or a translator. Ask a Hollywood story boarder or a San Francisco cabbie who’s competing with the driverless taxis recently unveiled there. All of them can tell you about how their jobs are changing. The questions left to us now are: what will we do about it? How will we make sure that people are okay?

READ: Why this Canadian author wants AI companies to pay up

In 2019, I began writing a novel about a poet who agrees to collaborate with a big tech company’s new poetry-writing AI. I already understood that the machines were coming for our jobs; I wondered what it would mean if they came for our art. My fictional poet, Marian Ffarmer, is a 75-year-old woman famous for her wit and her stanzas and her signature tricorne hat. The AI, Charlotte, is a potentially sentient computer program, just 39 days old. Marian agrees to the collaboration because Charlotte’s creators offer her $65,000—she wants the money because her middle-aged son cannot afford to buy his first home. 

Marian accepts a first-class plane ticket to California, sweeps into an office in Silicon Valley and takes a seat in front of a gleaming screen. “Hello,” she types, and Charlotte answers. I was inspired by the real-life story of Marianne Moore, one of America’s great modernist poets, who wore her own tricorne cap and cracked wise on The Tonight Show. In 1955, Ford Motor Company wrote to Moore, seeking her help to name their upcoming sedan. Moore could have said no. She was famous and esteemed, and Ford wasn’t offering any money. But instead, the 67-year-old, in all her gravitas, said yes. The Ford Thundercrester, she proposed. The Ford Silver Star. And stranger ones: Utopian Turtletop. Pastelogram. Mongoose Civique. The Ford execs lapped them up, sent her flowers, but in the end, they went with their own idea, the Edsel. It flopped.

Why did Moore do it? Perhaps she was tempted by the challenge—the invitation was too tantalizing. Never mind whether she was trading her talent, and arguably her self-respect, to a company’s marketing department. The secret truth is that art often feels cheap, even to artists. You think, “Might I not trade this away for treasure?” It’s not that different from the small surrenders we make every day, trading our personal data, our attention, for some free convenience or small delight online.

I was interested in that act of capitulation: of giving in to the machine, even when you know it probably isn’t good for you. Will AI really come for our art, or will artists like Marian relinquish it themselves, preferring the convenience of outsourced inspiration?

READ: How can we tell whether content is made by AI or a human? Label it.

The only way to know was to try the thing myself: to make my book, like Marian’s AI-assisted poem, a collaboration. When I started working on the novel that would result, Do You Remember Being Born?, ChatGPT was still three years from being deployed. But the company that developed it, OpenAI, had given researchers access to GPT-2, an earlier generation of the same eerie software, which predicts the next word in a sentence. I applied for access, and over the seasons to come, I scattered the novel with phrases generated by its successors.

I also wanted to use AI to write the poems generated by my fictional AI, Charlotte, in the book. But GPT-2, GPT-3 and even GPT-4 are curiously clumsy when composing free verse. A mistake was made in their training phrase, preventing the software from reliably formulating any poetry besides doggerel. For example, here is the first line of an imaginary poem:

When the memory struck, I was sitting by the window

If I feed this into GPT-4 with instructions for it to continue the poem with unrhymed lines “in the style of John Ashbery,” or almost any other contemporary poet, it offers something like the following (the bolded text is AI-generated):

When the memory struck, I was sitting by the window
A sliver of sunlight tracing the dust in its glow.
Whispers of yesteryears, like leaves in the wind,
Fluttered through chambers, the past rebeginned.

I struggled with this for months, trying to find ways to generate poetry in a consistent, contemporary style that matched the voice I had in mind for Charlotte. Finally, I requested funding from the Access Copyright Foundation and the Conseil des arts et des lettres du Québec, which allowed me to hire a consultant, Katie O’Nell (then a master’s student at Oxford). Katie and I trained our own custom AI, dubbed Moorebot, by taking an off-the-shelf system and fine-tuning it on a narrow collection of works, mostly poems by Marianne Moore. From examples like the line above, it produced lines like these (which I gradually edited together): 

When the memory struck, I was sitting by the window
I didn’t know I had become a bell
There are women who give up
Like the abandonment of a mineral inheritance—
a gypsum mine, a limestone
quarry. I remember being chipped
by a pickaxe
but now, as I say
I am a bell.

These words would not be worthy of Moore. But they suited the book and its synthetic poet.   

At first, the most surprising thing was that AI responses could thrill me at all: a turn of phrase, a ribbon of metaphor, that felt like an act of magic. Most of today’s tools, like ChatGPT, have been deliberately constrained and calibrated to make them boring, placid, reliable. But the technology undergirding the software has no such scruples, and can be configured to exhibit what almost seems like creativity. (Those who pay a fee for full ChatGPT access can find, deep in its settings, a variable called “temperature.” Turn the dial up, and the system’s choices will literally— statistically!—become more unpredictable.) 

As time went on, I also became fascinated by the systems’ persistent, sometimes arbitrary, blind spots—that difficulty with unrhymed free verse, for example, or its attraction to certain clichés. 

“Echo,” I called out into the room. “Echo my echo.”
The sound of my voice bounced off the walls.
The sound bounced back to me.
The sound of my voice came back to me.
The sound of my voice bounced off the walls.
The sound bounced off the walls.
The sound of my voice was swallowed up by the space.
The sound ricocheted like a transparent bullet off the walls.

Making literature with an AI felt like collaborating with a slot machine: pulling away at the handle, awaiting a row of cherries. It could feel lazy, yes, and at times addictive, yet I couldn’t shake the recognition that it had an inciting (insighting?) quality: that it felt generative in that other sense, sparking something in me, sending me down a new and untrammelled path, which was still my own to explore.

The roses’ color was
black and golden.
The roses’ color was
red.
The roses’ color was
affected by the blade.

Late in the process of writing this novel, I found myself with a dilemma: how ought it to end? Should Marian succeed, writing something magnificent with algorithmic aid? Should she fail, finding its “help” a form of sabotage? Neither seemed quite honest—and my own AI, to be clear, couldn’t solve the dilemma either. The solution I came to over time belongs in the closing pages of the book, for the reader to discover, but it has to do with recontextualizing the problem: people are not okay. Too often, life under capitalism is a zero-sum game. A competitor, a computer, a sentient poetry-penning robot—any of these could come and take your job, whisking the money from your future into theirs. 

But creativity is not a zero-sum game, and art needn’t be. The exchange of ideas, the play of inspiration, doesn’t come at the expense of anyone. The only idea that is discredited by artistic collaboration—with an AI, with anyone—is the myth that creation should be hermitic. In fact, we’re nourished by our muses, our mentors, our students, our peers, our children, our lovers and even by our tools: our libraries, our pigments, our software. AI is not that different from photography or the internet. The history of art is a history of interconnected influence, and artists are resilient. What makes us compete is the market, and even if we must visit that place, we are not obliged to live inside it. 

The machines may be coming for our jobs, it’s true. But nothing can come for our art. It’s a commonwealth—it isn’t anyone’s to take.

---

Sean Michaels is a Giller Prize–winning author whose work has appeared in the New Yorker, the Guardian and Rolling Stone. His most recent novel is Do You Remember Being Born?