There's considerable excitement about generative AI in humanities academia at the moment, with the launching of a new Duke University Press-hosted journal, Critical AI, and a new dedicated grant from the NEH on Humanities Research Centers on Artificial Intelligence. In some ways, the buzz over generative AI seems comparable to the buzz a decade ago over Digital Humanities, which brought fresh ideas and methods to the humanities, and seemed to be growing apace (though mostly outside of core humanities departments). That said, it remained somewhat of a niche activity with a certain ceiling of engagement.
The engagement with generative AI seems to be different. For one thing, the technical challenges involved in doing text analysis or computational stylistics are not present with generative AI; anyone can use ChatGPT or Gemini, and millions of people already are. Second, the utility of the technology is obvious and immediate (even if it too might be overstated in certain ways, both positively and negatively). To my mind, it seems important for humanities scholars, writers, and artists to engage with generative AI constructively but also skeptically. What are some possible ways the technology could assist us with existing research problems and questions? What are some new questions or problems it might allow us to ask?
Here, I'll take on just one of the many questions I've been thinking about with respect to generative AI: What might generative AI mean for the future of writing? If generative AI works as its supporters are advertising (or comes to work that way soon, as the available products proliferate and are continually refined), won't creative writers across a broad spectrum of creative areas be inclined to use it, perhaps not to replace creative labor, but at least to augment it in certain ways?
Along these lines group of writers have been thoughtfully experimenting with generative AI to achieve particular effects. In a class on representing AI I'm teaching this spring, I recently assigned a trio of published texts that do some version of this, Stephen Marche's Death of an Author, the collection of poems called I Am Code, and Sean Michaels' novel Do You Remember Being Born? How successful are these experiments?
Both of the novels in the group are set in the present moment, and prominently feature tech companies as agents in the plot. In Michaels' novel, a celebrated elderly poet is hired by a tech startup on the verge of unveiling a powerful new AI companion to jointly author a long poem with a generative AI fine-tuned to her own poetic style. In Death of an Author, we meet an eccentric billionaire who aims to create an AI replica of a celebrated novelist named Peggy Firmin. This commonality is of course not an accident -- and indeed, much of the coverage of the current generative AI boom is linked in some way to its corporate culture and its charismatic executives.
Death of an Author is described by its developer/author Stephen Marche as containing text that is 95% generated by a variety of generative AI platforms, including ChatGPT, Cohere, and Sudowrite. The choice of a murder mystery is an obvious one for a first experiment with generative AI -- detective fiction is well-known as being structured around strong and identifiable genre constraints, even as well-known writers in the genre often deviate from those constraints.
The novel as written is replete with literary and theoretical Easter Eggs, from the title (clearly a reference to Roland Barthes' Death of an Author), to the many references to writers like Edgar Allen Poe and Arthur Conan Doyle one finds in the story (the protagonist in the novel is named Gus Dupin [after the detective in Poe], who is the author of a work of literary criticism called The Purloined Author [another reference to Poe], and the plot as a whole borrows its essential skeleton from Doyle's "The Problem of Thor Bridge")
One of the limitations in the believability of the story in Death of an Author is its overly optimistic sense that AI replicas can be even remotely equivalent to the actual persona and being of a human being (in the story, Peggy Firmin first agrees to be replicated, but then has second thoughts about it). If this is ever going to be possible, it certainly isn't on the horizon with anything like the large language model-based, statistics-driven generative AI that is currently in use. (Which isn't to say that people aren't going to try. Internet sex workers are already selling AI bot versions of themselves to subscribers, and one can easily imagine a scenario where people with terminal illnesses sign up to have their speech and image captured for digital avatars that might persist after their death. One thinks of Black Mirror's "Be Right Back" episode...)
That said, there are some intriguing ideas in Death of an Author as well, gesturing toward how AI might transform fictive universes. Here, for instance, is the character Peggy Firmin in an interview from the text of the novel:
We're going to see interactive mystery bots that will allow users to solve puzzles and uncover secrets. We'll also see stories created specifically for individuals inside their experience, the ability to recreate dead relatives through AI technology. Stories where the audience doesn't even know they're stories. character who are so deeply felt that they aren't characters at all, but you become the character. It's going to be a gorgeous mess.
In effect, what Peggy Firmin is imagining here is something roughly akin to the gamification of works of fiction (which in point of fact is not so revolutionary to anyone who has played narrative-driven video games like The Last of Us). In Death of an Author, several 'characters' in the novel who only manifest via telephone or email are revealed late in the text to be bots, who have drawn the protagonist Gus Dupin into an "augmented narrative reality" mystery regarding the death of the 'author' indicated by the title.
While Death of an Author is a fine first experimental attempt at an AI-generated novel, there are a number of moments where the text seems to drift slightly off course, and the story as a whole feels strangely flat and affectless. Indeed, the best-written part of the book is probably the human-authored Afterword, where Marche makes the case for engaging in the experiment:
So little of how we talk about AI actually comes from the experience of using it. ....Like the camera, the full consequences of this technology will be worked out over a great deal of time by a great number of talents responding to a great number of developments. But at the time of writing, almost all of the conversation surrounding generative AI is imaginary, not rooted in the use of the tool but in extrapolated visions.
I tend to find this persuasive -- as I have been exploring different AI platforms and trying my own experiments in recent weeks, I've been pleasantly surprised by some of the results. We may find ways to make generative AI useful for certain specific tasks without losing what we really value in human-created art.
Another hint Marche gives relates to the possible value of generative AI in helping authors find 'heteroglossia':
I found the transformer-based AI shockingly good at what the narrative theorist Mikhail Bakhtin called heteroglossia--the novel's ability to incorporate other forms of discourse inside itself. If you ask linguistic AI to imitate a mode of speech, any mode, it can do so to an uncanny degree.
This strikes one as being potentially useful for writers who know their own voices and are comfortable in their own idiolects, but who might struggle to render compelling accounts of the voices of others.
Marche elsewhere describes the process by which he created Death of an Author using a trio of generative AI tools. However, one especially important part of the process apparently happened offline: "I worked out the plot during a long skate with my daughter and a walk with my son (better techniques than any machine)." While 95% of the text of the novel as published was generated by an AI, if the essential plot points, characters, and narrative framework were all generated by Marche it seems hard to see whether the AI really played a positive role in the novel's composition -- or whether it was, in effect, a bit of a gimmick.
Far and away, the most compelling experiment of the three might be Sean Michaels' Do You Remember Being Born? However, by contrast to the other examples mentioned, it's probably worth noting that this is also the text that had by far the least proportion of text by generative AI.
Michaels did use a specialized version of ChatGPT weighted towards the poetry of Marianne Moore and a small array of contemporary poets to generate the poetry included in the text. While the author referred to this version of the chatbot as a 'MooreBot,' fans of Marianne Moore might observe that the actual output is not that much like Moore's poetry (the vocabulary is more constrained, and punctuation is less robust, among other things). As intriguing as some of the outputs are, they constitute a fairly small segment of the overall text.
All of the most compelling elements of the story are human-authored -- the idea of an aging poet who has achieved success, trying to learn to write in collaboration with a machine; the interest in the experience of being a parent who has had to make sacrifices in the interest of also pursuing a career as a writer; and the attempt to stick closely to one's principles at a time when tech-dominated capitalism seems to rule the world.
An especially salient passage from near the end of the book might be this one:
“My whole life I had believed that understanding myself required me to keep others at a distance, lined up on the far side of a river. That evening of counting I had not felt so certain. That evening I felt like a room with doors open, for others to explore, and that from their explorations I could start to ascertain my shape. We are not the people we think; we cannot really see who we are. Here, on Sunday in San Francisco, I had the same impression that I might unfasten the locks and lower the drawbridge, that I might not be a fortress but a space for others to pass through.”
Here, Marian Ffarmer is describing how she found a process that would actually work for her -- she found a younger poet to be a co-author and interlocutor, alongside the generative AI that had been trained on her own published writings (but which, she's discovered, is actually an entity very different from herself). Together the two humans and one machine produced a finished text that is described in Michaels' novel, but not shown directly.
To my eye, the intriguing possibility is the idea of writing where the author has fewer controls over self vs. not self in the process of creation. As we learn to write differently, perhaps with generative AI tools assisting us, we might have to imagine the writing self, as Michaels says,"not [as] a fortress but a space for others to pass through."