Description
Accepted paper: 'The AI Question'Since classical antiquity, there has been long-standing interest in the ‘Homeric Question’. This question covers several topics, most of which are related to the nature of authorship of the Iliad and the Odyssey. The Homeric Question focuses on who, exactly, Homer was, whether some of the most well-known epic poems of all time were the result of collaborative or single authorship, and under what circumstances the poems were composed and subsequently written down.
The same questions can plausibly be asked of the works generated today by Artificial Intelligence (AI). Whether we are talking about a chapter inspired by the Harry Potter franchise, or the output of neural networks and text-generating bots, it remains unclear as to who, exactly, ‘wrote’ such works, whether they were composed by individuals or collectives, and indeed how such texts are produced in the first place. The similarities continue; just as the Homeric question focuses on the seismic developments surrounding the transition from oral to written culture, so AI-generated works ask us to reflect on what the transition from written to automated culture will involve, along with what impact this may have on the nature and reception of literature.
My short paper will focus on this so-called ‘AI Question’. I will look at four examples of AI-powered authoring tools, from the simplest to the most advanced, and use the output of these tools to explore the way in which AI is contributing to a new form of literary culture. The aim is to consider how the questions asked of Homer can enhance our knowledge of AI-generated literature, as well as how AI-generated literature might in turn shed light on important transitions in the ancient world.
This paper will begin by looking at the AI-powered keyboards created by Botnik. These keyboards can be pointed at any body of literature and will then suggest the next most likely word. In this paper, I will share the results of my experiments pointing Botnik’s keyboard at the works of Homer. As I will show, the simple Markov Chains behind the keyboard require a significant level of human input to generate epic poetry. In fact, it is virtually impossible to write epic poetry using these keyboards. Instead, the most logical outcome is to write absurdist poetry. Moving on to the neural network GPT-2, I then show how more advanced AI-tools can be tasked with suggesting entire phrases based on the first line of the Odyssey. While human choice is still important in determining which phrase is the most appropriate (or absurd), GPT-2 provides significantly greater input based on its training data, which includes a cross section of human literary output. Following GPT-2, I turn to Sudowrite, powered by the latest neural network GPT-3. Using Sudowrite, it is possible to ask the AI to generate an entire page of epic poetry based on the first line of the Odyssey. GPT-3’s output is perhaps the most coherent of the three, though appears prone to extensive plagiarism. However, it demonstrates the potential for AI to co-write anything from news reports to creative literature. Finally, I turn to InspiroBot and AI-generated inspirational phrases. These phrases come closest to the fantasy of AI authorship, not only in the sense that an AI appears to be writing independently of human input (all the human does is click a button), but also that the AI behind InspiroBot seems to understand motivational posters and quotes enough to be able to spoof them.
Using these examples, I will argue that AI’s do not ‘author’ works so much provide endless variability inspired by existing formulas and ideas. The most creative work emerges at the intersection of human writers (or editors), and AI tools. By determining the contexts of a work, AI-powered writing tools are reminiscent of historical examples, such as the composition of the Homeric poems, when the possibility of a single, stable epic was far from certain, and when each performance remained in flux, caught between an individual’s creative input and a dataset of myths and formulations.
| Period | 19 Oct 2022 → 20 Oct 2022 |
|---|---|
| Event type | Conference |
| Degree of Recognition | International |