The first time I played the tabletop game Fiasco, it wasn’t the story my friends and I made that blew me away. It was the realization that I had just experienced the limitless possibilities of collaborative writing, that the novels I loved featured just one way their narratives could have played out. Alice could have transformed the Mad Tea Party into Wonderland’s first organic tea shop. Don Quixote could have devolved into a windmill-killer for hire.
Later I realized the similarities between tabletop games and ways novelists challenge their narrative choices, from literary constraints to automatic writing to William Burroughs’ cutup method. Sometimes authors even bring in collaborators, be they humans or machines.
The newest collaborative machine that has everyone talking is, of course, ChatGPT, a large language model (LLM) often called artificial intelligence, although its cognizance is nonexistent. Working with ChatGPT, an author inputs a prompt request and can choose from an infinite number of outputs. If you don’t like a paragraph ChatGPT wrote, you can edit your prompt and ask it for another one. These tools can guide anything from character names to plot points. “Alright, ChatGPT, that character in my sci-fi story is a member of the Merry Men? OK, let’s try Robin Hood in outer space.”
Death of an Author, by Stephen Marche, is the best example yet of the great writing that can be done with an LLM like ChatGPT. Not only is it an exciting read, it’s clearly the product of an astute author and a machine with the equivalent of a million PhDs in genre fiction. ChatGPT read basically the entire internet and all of literature, finding billions of parameters that go into “good” writing.
At the level of narrative, Death of an Author reads like your everyday detective novel, following a scholar of crime and cyber fiction named Augustus Dupin as he races to clear his name while uncovering the greatest mystery of his life: Who killed his favorite author, Peggy Fermin? An
AI? A human? An AI posing as a human? And why?
The story is full of holographic avatars and deceitful tech CEOs, as well as everything you might expect from crime fiction: mysterious letters, hidden codes, nosey journalists, and gruff cops. Clever Easter eggs are hidden throughout the text, referencing the history of mystery novels. The protagonist’s name, for example, is a nod to Edgar Allan Poe’s C. Auguste Dupin, while the dead author’s funeral reads a lot like the beginning of And Then There Were None.
Death of an Author doesn’t try to hide many of the quirks that come from collaborating with ChatGPT, which favors Thomas Pynchon-esque monikers, like a literary agent named Beverly Bookman, and fake book titles, such as God, Inc. and Tropic of Tundra.
Descriptions also bear an alien, computer imprint in their discordant and nearly nonsensical images. One character walks away “like a record being put back in its sleeve.” Workers at a tech company are compared to “digital mountaineers scaling the face of God for a sprig of robotic edelweiss.” Multiple characters are described as “glowing,” but a fictional version of author Michael Ondaatje gets the most luminous treatment. At one point his “hair was a glistening white, and it radiated a soft aura around him” and later his hair is “radiating a soft aura.”
The LLM voice is also apparent in several out-of-character moments. Augustus, the staid, academic protagonist, refers to his estranged spouse as “wifey.” At another point, he seems unaware of one of the biggest biographical facts of the murdered author’s life, even though he’d been studying her work for years. These slippages are byproducts not only of the LLM’s idiosyncrasies but of a human author’s process in composing a novel with LLMs the only way currently possible: piecemeal.
An LLM cannot write a cohesive, novel-length narrative from start to finish. At present, ChatGPT can produce roughly 600 words at a time, so in order to complete a novel, a human has to feed it prompts and then collage its outputs into a complete story. One prompt might be something like “Describe the death of an author in the style of CBC news.” The next might be “Write Augustus’ response to this death.” The computer can’t keep track of the minutiae of plot and character, leaving holes in the process.
In his afterword to Death of an Author (required reading for anyone who wants to think seriously about the future of LLM-assisted writing), Marche explains this patchwork process of composing the novel. He reread some of the great detective fiction writers, like Agatha Christie, Raymond Chandler, and James Myers Thompson, and employed ChatGPT to produce passages in their styles.
To polish the outputs into more readable language, he ran the text through Sudowrite, another LLM that allows for more stylistic authorial control (making sentences longer or shorter, rephrasing text, etc.) and then used yet another program, Cohere, to generate poetic similes, refining the language even further. Marche’s goal might have been Chandler, but to this reader’s ear, the prose is closer to Dan Brown—compulsively readable, but not in danger of winning the Edgar.
Despite the use of so many different programs and styles, this text has Stephen Marche’s signature all over it.
Marche even said as much to The New York Times: “I am the creator of this work, 100 percent.” What’s striking, though, is what he says next: “But on the other hand, I didn’t create the words.”
It’s important to meditate on this renunciation of authorship because it seems central to current misunderstandings about what LLMs are and why they make us so nervous. On the one hand, Marche acknowledges his role in the creation of the text: “I had an elaborate plan … I have a familiarity with the technology … I know what good writing looks like.” On the other, he chose to publish the work under a pseudonym, Aidan Marchine, a portmanteau of machine and Marche. Calling him the author of Death of an Author wouldn’t just be a tongue twister, it would be, according to Marche, “inaccurate as a matter of fact.”
This seems like a missed opportunity. As someone who has just finished writing two novels that incorporate LLM language, I agree with Marche that only a good writer will make anything worthwhile with these programs. Because of this, it seems important to acknowledge the human hand in every aspect of the writing process.
Even the choice to include a particular LLM output over another is a human decision, not unlike the selective reframing employed by artists like Marcel Duchamp and Andy Warhol. Giving creative credit to the LLM seems to turn an elaborate collaboration between a human and a machine into a flashy tech gimmick. And it plays into the hands of forces who claim that writing with LLMs is not “real” writing, nor is it worthy of copyright protection, as the US Copyright Office recently argued.
LLMs are not authors, nor do they possess intelligence. They are simply tools. They can’t produce text at a random moment out of their own creative inspiration.
They are computer programs trained to recognize patterns in how we write and then use these patterns to produce language that appears like conscious and coherent thought. At least for the foreseeable future, these programs do not operate without prompts. They can’t produce text at a random moment out of their own creative inspiration. They can’t prompt themselves to fulfill apocalyptic fantasies and take over the planet. They begin and end with human direction. As such, the material that LLMs produce should be seen as a collaboration between a human author and a machine. The author asks the machine for language and then creatively determines what to do with the machine’s outputs.
It might be helpful to situate Death of an Author not in the tradition of LLM writing, but in the larger field of literary supercuts, or works of fiction made entirely out of found language. While the history of literary supercuts is less known, fiction writers have been incorporating found language for centuries. Al-Jāḥiẓ, a Medieval Arabic author, borrowed plenty from other sources. Herman Melville’s Moby-Dick begins with 13 pages of found whale descriptions.
Literary supercuts have appeared as recently as Graham Rawle’s Woman’s World (made entirely out of language from 1950s women’s magazines) and Kathryn Scanlan’s Kick the Latch (a collage of interviews with a female horse trainer). Many of these supercuts, like LLMs, look for patterns in found language and use them to cluster fragments of text into narratives that reveal something new and often startling about their sources.
I did something like this in my recent novel, The Nature Book, which is made entirely of found nature descriptions from 300 novels. To write it, I operated like a data scientist or human algorithm, studying thousands of pages of nature descriptions, looking for patterns, and then using them to guide the narrative.
Death of an Author seems to function in a similar fashion. After studying patterns in how crime fiction works, Marche gathered them into a novel outline. He then took the LLM’s outputs (themselves the product of machines studying patterns in writing) and pieced them together into a final form. This process is not so far removed from a Burroughsian cutup, but it veers from Burroughs’ opaque writing and toward the seamlessness of the best genre fiction.
Like supercuts, what gives Death of an Author its conceptual rigor are the rules and guidelines Marche outlined for the text. He decided that 95 percent of the novel would be computer-generated, with only minor derivations from the LLMs’s output; that it would be novella-length; and that the text had to be “compulsively readable, a page-turner.”