In his afterword to the short murder mystery Death of an Author, the writer Stephen Marche invokes a concept called Moravec’s paradox. Hans Moravec, a robotics scientist, observed that tasks human beings find challenging, such as playing chess, are easy for computers, while many of the actions human beings effortlessly perform without conscious thought, such as perception or oriented movement through space, are extremely difficult for the machines.
Moravec’s paradox is a useful way to think about the surprising ways that Death of an Author, described by its publisher as a “groundbreaking experiment” in artificial intelligence, succeeds. Jacob Weisberg, the head of podcast production company Pushkin Industries (and a former Slate editor in chief), asked Marche, a journalist who writes about artificial intelligence, to make Death of an Author earlier this year. The goal was a novella whose text was to be 95 percent computer-generated. Engineering such a text is not as easy as typing “write a murder mystery about the death of a famous writer” in the ChatGPT prompt box. Instead, Marche describes a painstakingly piecemeal effort, requesting specific passages written in particular styles, tinkering with them, and then using them to flesh out a skeleton plot of his own devising.
Death of an Author
By “Aidan Marchine.” Read by Edoardo Ballerini. Pushkin Industries.
Slate receives a commission when you purchase items using the links on this page. Thank you for your support.
The result has, as Marche puts it, “something alien” to it. Death of an Author is suffused with an eerie placidity that prevails even when something eventful or alarming is happening in the story. But in the case of this narrative, at least, that effect feels appropriate. Death of an Author begins with the main character, an academic named Augustus Dupin (the first fictional detective story, written by Edgar Allan Poe, featured a detective named C. Auguste Dupin), learning that Peggy Firmin, the famous writer whose work he studies, has been murdered. To his surprise, Dupin receives an invitation to Firmin’s funeral, despite the fact that they have never met. The ever-bewildered Dupin learns that the police have clues implicating him in the murder, and he receives an accusatory call from the victim’s daughter. As a scholar of crime fiction, Dupin knows that in a murder mystery, “the murderer was the only one who wasn’t suspicious,” which puts him in the crosshairs. Concerned, he has begun to investigate the crime himself when he receives by mail a short story sent by Firmin in which a fictional writer’s death and funeral exactly resemble her own.
All of this is very meta—reminiscent not only of classic mystery fiction but of certain literary deployments of crime-fiction tropes for metaphysical purposes. So it’s no surprise to learn from Marche’s afterword that in preparing to produce Death of an Author, he studied up not just on Raymond Chandler and Agatha Christie but also on Paul Auster, Haruki Murakami, and Carlo Emilio Gadda. (In fact, I laughed out loud when the novel paused to describe in detail how Dupin prepared himself a dinner of fried mushrooms on toast. Interludes featuring methodical accounts of cooking are a Murakami signature.) Like Death of an Author, these human-written, cerebral, quasi-crime novels tend to have a muted emotional range, as if a cool, damp sheet has been laid over the proceedings.
Nevertheless, Death of an Author is more intriguing than many of the human-written mysteries I dip into in the course of my work. The prose style is better than average. When Dupin receives an email of condolence from his department head after Firmin’s death, the cut-and-paste wording of the note reminds him that she “relished the cruelties of boilerplate.” Marche explains that he loaded a bunch of descriptions of the smell of coffee into a natural language processor called Cohere, then asked it to output similes until he got one he liked: “The smell of coffee was like a fog burning off a field,” which is striking, even if I’m not entirely sure that dissipating fog has a smell.
True, if you simply ask, say, ChatGPT to write a story about your cat in the style of Vladimir Nabokov, the results will be disappointing. But Marche insists that “if you make bad art with a new tool, you just haven’t figured out how to use the tool yet.” For some reason, the worst course is to straightforwardly ask the tools to imitate Raymond Chandler. Instead, Marche writes, “what you need is to have it write something about a murder scene in the style of Chinese nature poetry, then make it active, then make it conversational, then Select All and put it in the style of Ernest Hemingway. That gets you something interesting.”
Where the A.I.s Marche used were hopeless, it turned out, was in coming up with a plot. Even the most elementary of stories—one thing happens, which causes another to happen, then another, in any remotely interesting sequence—was beyond their grasp. Marche had to work all of that out in advance. This is where he brings up Moravec’s paradox: “What James Joyce did in Ulysses, bring a whole city’s voices together in a vast syncretic amalgamation—no problem,” he writes. “Come up with a simple story about something that happened that’s interesting—quite tricky.”
This is a complete reversal of how writers came to value their work during the 20th century. The British novelist Martin Amis, interviewed by the Paris Review in 1998, complained of the basic prose produced by some of his contemporaries. Style—an intensely worked, attention-getting way with sentences—was Amis’ thing, both as reader and writer. “If the prose isn’t there,” he told his interlocutor, “then you’re reduced to what are merely secondary interests, like story, plot, characterization, psychological insight and form.” This position is extreme—enough so to suggest that Amis was trolling, but that doesn’t make him an outlier. Anyone who read her fair share of late 20th-century literary novels found many to be packed with beautiful sentences but hopeless at making readers eager to know what happens next.
Marche implies that writing ambitious or elegant prose is easy for A.I.s but difficult for humans, and vice versa when it comes to devising engaging plots. But are good plots really so easy to invent as to warrant being dismissed as of “secondary interest” in judging a writer’s work? Most humans can tell a coherent story, it’s true, and most of us can’t come up with a breathtaking metaphor. On the other hand, if creating good plots is so easy, then how come brilliant plotters like Christie—or even just competent ones like James Patterson—aren’t more common? If they were, Christie wouldn’t be the bestselling author of all time.
It’s not hard to see why A.I. finds plot more difficult than whipping up a pleasing prose style. Playing around with words until you produce an arrangement that pleases your discerning human partner is basically a numbers game. But coming up with a decent murder-mystery plot requires misdirection. Christie’s stories often hinge on how easily our expectations can be used to deceive us. We glimpse the back of a redheaded woman wearing Cynthia’s clothes and hat, and we assume that it’s Cynthia, even when we haven’t seen her face. We know how a particular type of detective story is supposed to work and don’t consider that the author might break its conventions. Such deceptions require a theory of mind, the ability to anticipate what a reader is thinking and how she can be tricked. No A.I. can do that—at least not yet.
More than one reader has told me that they don’t want to read another writer’s work, no matter how compelling the story, unless it “puts pressure on the language” or uses words in some unprecedented or remarkable way. But Death of an Author suggests that nonhuman intelligence is well on its way to being able to independently craft what Marche describes as “elaborations of the most complex style,” or passages of genuine beauty. (For now, the programs require a human editor to pick the best similes, but perhaps in time A.I. can “learn” an editor’s taste?) What’s typically viewed as the lowest form of literary achievement, however—elementary plotting—is the written equivalent of walking down a busy sidewalk, the kind of activity we think of as easy only because it comes naturally to us. Yet this activity deploys so many complex systems that it can utterly stump a machine capable of solving higher mathematics problems in milliseconds.
Death of an Author will be available as an e-book, but Pushkin Industries is primarily an audio production company, so the audiobook version ought to be regarded as its definitive form. Marche writes that Weisberg hoped to use an A.I. narrator, but none of the currently available synthetic voices struck either Marche or Weisberg as tolerable. So they hired Edoardo Ballerini (profiled in the New York Times as “the voice of God”) to perform the book. Though Weisberg told the Times that the technology is moving so fast that he thinks the A.I. voices that weren’t good enough six weeks ago are now “up to snuff,” Marche isn’t convinced. “You know who doesn’t have to worry about losing his job to AI?” Marche writes. “Edoardo Ballerini.” With some skills, it seems, the human touch is essential.
Plot-making appears to be one of them, and those writers who excel at it, masters like Agatha Christie or Stephen King, are to run-of-the-mill storytelling what ballet dancers are to the rest of us ordinary shamblers. Perhaps they will get more credit for this from literary critics in a future in which lovely sentences and poetic similes can be mass-produced.