I want AI to tell me how the story ends

We may be surprised and delighted if we ask computers to fill in the blanks of unfinished novels.

Richard Yates, whose body of work includes a novel he did not finish before his death in 1992.handout

I have spent the last month reading “Uncertain Times,” a novel by Richard Yates left unfinished at his death in 1992. It follows William Grove, a struggling speechwriter with a penchant for drink.

Discovered in the author’s freezer, the pages feature Yates’s trademark dissection of middle class ennui. The story is romantic, muddled, and tragic. More than that, it offers personal insight into the trying atmosphere of Robert F. Kennedy’s office, where Yates had worked. Sadly, it has yet to be published, languishing in the archives of the Boston University library, where I went looking for it.

Accompanying the manuscript are myriad notes. Scanning them for clues about the fate the author envisioned for Grove, I was curious about the potential of generative AI to flesh out the story. The critic Frank Kermode wrote that readers look for “the sense of an ending” — a resolution that bestows meaning on all that came before it. But with an unfinished work, this need goes unmet, which is why it’s so hard to accept the novel as it is: fragmented and imperfect.

And that’s why I want a new AI dream machine to finish the novel. By synthesizing the notes and the rest of Yates’s fiction, AI might someday approximate the Yatesian voice better than a human could. Perhaps a chatbot could realize the author’s vision — or at least come close.

For now at least, this isn’t possible on the freely available chatbots GPT-4 and Sudowrite. It’s not clear how much of Yates is in their training data, and I am no expert at writing prompts that would generate a decent output. But AI systems are improving rapidly. Anthropic, developed by a San Francisco startup, is capable of analyzing book-length manuscripts in under a minute. It seems inevitable that with sophisticated and learned prompts, chatbots will soon be able to unearth text that approximates Yates’s exquisite craftsmanship and melancholy.

There is an entire realm of unfinished novels that could be completed. We could discover what happens to Jane Austen’s Charlotte Heywood in “Sanditon” or how Cecilia Brady copes with her grief in F. Scott Fitzgerald’s “The Last Tycoon.” Then there’s Kafka’s “Amerika,” Nabokov’s “The Original of Laura,” and the mystery of Sylvia Plath’s second novel, “Double Exposure.” And if we go back to the earliest texts in Greek and Roman literature, we might even complete Euripides’s “Iphigenia in Aulis” and Plautus’s “Amphitruo.” Many of these works are in the public domain.

Using AI to complete unfinished novels has a kind of precedent. Long before the emergence of chatbots, literary modernists like André Breton and Robert Desnos experimented with “automatic writing.” This approach often made use of prompts. The goal was for the unconscious mind to assume command of the creative process. Words streamed from beyond the writer’s awareness, as though emanating from a spirit.

The idea is that all creation is necessarily an act of synthesis; all text is nothing but sources from elsewhere. The English novelist Tom McCarthy has said: “I see what I’m doing as simply plugging literature into other literature. For me, that’s what literature’s always done. If Shakespeare finds a good speech in an older version of Macbeth or Pliny, he just rips it and mixes it. It’s like DJing.”

The most seductive thing about an unfinished work is the transfer of power from the author to the reader. Without the author, readers gain a sense of control. A partial manuscript holds the promise of revelation. Does it foreshadow the author’s death in some way? Does it answer philosophical questions about the ethics of posthumous publication? And more important: What would the author have done with those sex scenes? The reader is free to speculate.

So why would we want to finish unfinished novels with the help of AI? First of all: sheer pleasure. Readers would be able to dabble in automatic fan fiction, writing outrageous plots and bad dialogue and even placing themselves within the story. For would-be novelists, AI could be instructive, continuing the narrative through a dissection of the novel’s voice, language, themes, and setting. For scholars, completing another writer’s unfinished novel would present a compelling research exercise, one that offers potentially deeper insights into the author’s craft and intentions.

There are also commercial reasons. The estates of Ian Fleming, Agatha Christie, and V.C. Andrews have authorized “continuation novels” by new authors. The hope is to preserve the value of the literary franchise. Such continuations will soon be able to make use of AI.

The collaboration of human and machine does not necessarily mean plagiarism or the death of the author, charges that are often leveled at generative AI. Rather than an erasure of creativity, AI allows for new ways to engage with the novel form.

I want to read a finished copy of Yates’s “Uncertain Times.” Let’s see what the chatbots can do.

Nathan Dunne is the author of “Lichtenstein” and the editor of the essay collection “Tarkovsky.”