Alessia Giulimondi studied Classics at Sapienza University of Rome and Linguistics at Utrecht University. Since September 2026, she is now working on the epistemological obstacles in the history of artificial intelligence as part of Professor Yuk Hui’s NWO project Epistemologies and Political Economy of Automation. We welcome Alessia to the ESPhil community!
Alessia, your first thesis was on Roman history, and now you’re researching large language models. How did AI catch your interest?
The fil rouge connecting all my interests has always been language as a means to understand the human. In my thesis in Roman history, I examined the socio-cultural influence of the late Roman Empire on the Iranian empire of the third century CE by looking at recurrent themes in Middle Persian texts and inscriptions. I have learned that perceptions and understandings always leave their mark in the (un)intentional accumulation of linguistic traces and, by looking at the Romans through ancient Iranian texts, I saw how language shapes reality to the point of influencing the culture and politics of entire empires. Thus, when computers finally succeeded in handling language, I became interested in how computers can (re)shape our world just by joining the conversation. If language operates in the world as a creative tool to modify, build, disrupt or suspend reality, what happens when computers are capable of doing the same?
And when you study the history of AI now, do you get to draw on some of the skills you developed working with ancient texts?
I do! My background in Classics has equipped me with philological approaches that are useful for navigating the unfolding of Western thinking about intelligence and machines. Philology teaches us the study of documents from which we understand systems of thought and how they relate to each other. A text has a history of its translations, annotations and interpretations accumulated across decades or centuries. Philological tools uncover this hidden history, and this is what we need for an epistemological history of AI.
Your interest in language then took you from Classics in Rome to Linguistics in Utrecht. Does that training make you look at a chatbot like ChatGPT a little differently?
The current mainstream discourse about the question of intelligence and machines is driven by the simple fact that a machine is finally able to engage in conversations with humans. The advent of ChatGPT manifested the long-lasting intuition that intelligence is inextricably related to language. Linguistic fluency is the outward sign that agency and, most importantly, consciousness exist. Thus, language seems to stand at the centre of the question of intelligence. By following the way language has been modelled and (mis)understood from the beginning of AI history, we may find in it a profitable perspective to investigate the epistemological obstacles in the history of AI.
These obstacles often arise when an idea or explanation becomes so familiar that we stop questioning it, even when it starts getting in the way of understanding something new. Making progress can then mean rethinking what we thought we knew. How might that happen when we use AI to understand human language?
In language, and cognition more generally, epistemological obstacles are often found in the relation between our linguistic models and real human language. For example, one of the biggest puzzles in language studies is how children learn language so fast and so proficiently. Several models proliferate in linguistics, and LLMs have naturally come to the centre of this debate with several researchers using them as explanatory devices for the way children acquire language. That is, if LLMs learn language just by recognizing patterns in great quantities of data, with no prior linguistic architecture, we should assume that this is also what children are doing. However, before making such claims, we should ask ourselves how much our models can explain the cognitive mechanisms that underlie human language and in which relation LLMs and human language stand. Can machines be an obstacle to our understanding of language and intelligence? Or are they, perhaps, enhancing our knowledge by virtue of them being an obstacle?