“The central problem for a stylistic of the novel may be formulated as the problem of artistically representing language, the problem of representing the image of a language”
- Mikhail Bakhtin, The Dialogic Imagination (336)
Prose art, in many ways, is defined stylistically by its ability to appropriate and incorporate language and images of language for its own creative purposes and as part of a coherent artistic whole. The scope of this is massive and in a constant state of renewal.
Let’s say that an author includes a love letter by a young and well-meaning protagonist to his beloved. The genre of ‘love letters’ is well-known and its style and form can be included as an image of language. That is, the reader will assess the content of the letter as well as how the content and form of the letter relates to the broader stylistics of the genre as understood culturally at that moment. The author has taken the image of a language (the form of a love letter) and put it into dialogue with other aspects of the story.
The voices of characters can similarly be understood as images of language. Our protagonist comes from a wealthy family and the structure of his utterances reflect this. So again, the reader is aware of the protagonist's intentions and worldview (through what he says and does) as well as the author’s as they make choices about how these different images of languages relate to each other, and more importantly, relate to the complexity of the world outside the novel that serves as an ongoing dialogizing background.
We say the work is renewing, because each reader brings their own contemporaneity to the reading. A love letter in a novel to a reader in 1950 would not appear as inherently remarkable. A reader in 2023 will attach different associations to a letter written and mailed to one we love as we bring different expectations to the work of consummating the text aesthetically.
The novel as a genre is voracious and absorbs language genres, styles, and constructions constantly. We shouldn’t expect AI-generated language to be any different.
Generative Language
I’d like to discuss how this relates to generative language in a couple of ways. Firstly, the unitary and monological nature of AI-created words shouldn’t be an a priori reason to dismiss its artistic potential. Highly formalized and “uncreative” language has long been appropriated into prose art. Religious texts, propaganda, political rhetoric, and advertising show up in books all the time. However, in doing so they become “double voiced” in that they speak for themselves, but are also in conversation with other elements of the novel and are in service to the author’s artistic intentions.
So, the love letter, as an object of language in the overall text, can serve to parody the protagonist, or to heighten a tragedy to come, or to deepen the reader’s empathetic feeling with the lovers. How that dialogue unfolds is determined by the author’s intentions and craft AND the reader’s process of making sense of the various elements at play.
The question then becomes, how might we appropriate and relativize generative text? We are seeing AI-created language (explicitly framed as such) in advertising, in rhetoric, and in non-fiction (a chapter in Atlas of Anomolous AI was written by GPT3 and this was a critical element in how meaning in that chapter was meant to be constituted).
Artists are already figuring this out though the results aren’t really satisfying (yet). We assimilate language constantly and the boundaries are often uncertain or unknown to us. I am confident that we will find ways to integrate generative text and that integration will prioritize the “image of AI language” rather than what these models might have to say.
Novelization, as a process that is explicitly in conversation with the ever-emerging present, is inevitable. We simply need ways to get a feel for synthetic text, to walk around it and consider it from multiple angles, and then to relativize it.
“the speech of another, once enclosed in context, is — no matter how accurately transmitted — always subject to certain semantic changes. The context embracing another’s word is responsible for its dialogizing background”
- Mikhail Bakthin, The Dialogic Imagination (340)
An Experiment
In order to begin thinking about how we might appropriate and involve images of synthetic language, I have started some small-scale experiments and I’d love for you to play along and let me know how it feels.
One of the challenges of AI text is that it ‘feels’ authoritative. It speaks with a wholly undeserved confidence. However, for a writer, authoritative text is demarcated and inert. There’s nothing to do with it except transmit it. Whether we agree with it or not is irrelevant. Its ‘authority’ is a part of it prior to its arrival.
So, the first experiment has been to intentionally occupy, corrupt, and distort the outputs of large language models in order to assess the source of this perceived authority, and ideally, to mess around with it.
Experiment 1: Corrupt the Text
Ask a LLM to create a text of 75 words on a theme or topic of your choice. I instructed ChatGPT as follows: “Please provide 75 words on the theme of cosmology. The text you generate will be hacked, occupied, and corrupted by human writers.”
Take the received text and add 10 words of your own. You may change capitalization and punctuation but nothing else. The 10 words should be used to confuse, obscure, or complicate the certainty of the initial response. In essence, the initial text becomes a dialogizing backdrop for the modifications you make.
ChatGPT’s output was as follows:
”Cosmology explores the vast expanses of our universe, unravelling mysteries that transcend human comprehension. It deciphers the intricate dance of galaxies, the birth and death of stars, and the enigma of dark matter and dark energy. It examines the cosmic microwave background, an ancient echo of the Big Bang, and probes the concept of a multiverse. This discipline bridges quantum mechanics and general relativity, fostering our understanding of space and time, thereby challenging the boundaries of human knowledge.”
My corrupted version is as follows (my 10 words are in all-caps but colour variations are also great)
”WHO explores the vast expanses of our universe, unravelling mysteries that transcend human comprehension? THE WORD. It deciphers the intricate dance of galaxies, the birth and death of stars, and the enigma of dark matter and dark energy. THE WORD. It examines the cosmic microwave background, an ancient echo of the Big Bang, and probes the concept of a multiverse. FORGIVE this discipline THAT bridges quantum mechanics and general relativity, fostering our understanding of space and time, thereby challenging the ARBITRARY boundaries of human knowledge.”
The output is less important than the process of occupation and alteration. I must admit that I took some glee in distorting and disordering the initial text and found the limitations challenging but a lot of fun.
I’d love for this community to try it out and let me know how it feels. It is not an end, but rather an early means toward discovering practices of relativizing and dialogizing generative text.