02 - Why should I believe what you tell me is true?”: What Machine-Generated Homeric Poetry Tells Us about AI and Philology

Linguist Charles Fillmore once described a certain tension between two types of researchers in the field (Fillmore 1991: 35) - there are armchair linguists who note interesting facts about language and corpus linguists who count, rank, and quantify whatever they can in the language. The corpus linguist asks of the colleague: "Why should I think that what you tell me is true?" The colleague responds: "Why should I think that what you tell me is interesting?" In the emerging conversation about the role of AI in philological research, we can see the same tension between what is true and what is interesting arising in Classics.

How can results from black-box AI models provide evidence for philological and literary arguments? (e.g. Lau et al. 2018) To what level of technical understanding and ability should we hold computational linguistics and future classicists? What are the ethical considerations at play in the emerging field of Classics + AI? In this paper, I review work that brings together AI and ancient Greek poetry, including my current dissertation project in which I develop AI-driven computational approaches in order to model the "formulaic mechanics" of Homeric verse.

Consider, for example, the following lines of Homeric poetry:

αὐτὰρ ἐπεὶ δὴ δούρατ ἀλεύαντο μνηστήρωνἷξε τόθ ἔσαν οἶκον δὲ καλυψώ δῖα θεάων

But when they had avoided the spears of the suitors,
He came then to the house of Calypso, heavenly among goddesses.

The first line is from Homer (Od. 22.255), but the second is machine-generated. Classicists are easily able to identify that the second line is "fake" because, canonically, the suitors and Calypso never appear in the same context (a close reader will also notice that the second line includes among other things metrical errors). Although this line is not true Homeric poetry, its content is interesting. Both the suitors and Calypso represent obstacles to Odysseus' homecoming, and thus, semantically, there is an element of truth to the lines that this model produced. Even this one passage highlights the tension between the "true" and the "interesting," the original and the imitated, and the singular and the replicable.

Throughout this paper, I discuss how projects such as these balance the use of "black-box" algorithms with the evidence-driven requirements of classical philology through computational outputs like the analysis of metrical intertextuality (Sansom 2021) and experimental Homeric verse (Lamar & Chambers 2019) as an invitation to discussion within the field about ethical issues pertaining to authorship, researcher trustworthiness, and replicability in Classics.

---

Presenters

Annie K. Lamar, Stanford University

  SCS-9