Historians don’t know when the Ancient Roman text “Res Gestae Divi Augusti,” a chronicle of Emperor Augustus’s deeds, was first written, since these kind of epigraphs tend to not contain any written dates.
Enter our hero Aeneas — not the mythological forefather of Rome, but a generative AI model that’s been trained on Ancient Roman texts. According to The New York Times, the Aeneas AI pinpointed the date of the Augustus epigraph to around 15 CE, soon after his death in 14 CE.
Aeneas, developed by Google DeepMind, was able to do this because it mimics what historians do but faster: retrieving relevant contextual information, uncovering isolated texts, and analyzing them before arriving at a conclusion.
In a new paper in the science journal Nature, researchers from Google DeepMind and several European universities put the AI model through its paces and found that it was able to provide helpful information for historians in most cases, further cementing its apparent usefulness.
For history researchers, Aeneas is just one of a growing suite of AI tools that are helping them reveal more details about the ancient world.
“State-of-the-art generative models are now helping to turn epigraphy [study of ancient script] from a specialist discipline into a cutting-edge field of historical enquiry,” said study coauthor and classics and ancient history professor Alison Cooley, of Warwick University, in a statement about the research.
To study any type of ancient writing, historians have to comb through various archives all over the world while in pursuit of “parallels” — ancient texts with similar wording or from the same era in time — and compare them to the text they are studying. By using this method, researchers can extrapolate what the missing fragments may be saying or its context.
“Studying history through inscriptions is like solving a gigantic jigsaw puzzle,” study coauthor and University of Warwick historian Thea Sommerschield said at a press conference last week, as reported by NYT.
But this process is long and tedious, and requires historians to undergo years of specialized training.
In their evaluation of Aeneas, the study’s authors found that the tool was able to provide “useful research starting points in 90 percent of cases, improving their confidence in key tasks by 44 percent,” the paper reads. Furthermore, when human historians and the AI model worked in tandem, the results were even better compared to Aeneas or the historians working by themselves without any aid.
But what of Aeneas hallucinating fake results, which is a persistent problem with many AI models? The model gives probabilities on its predictions.
“Interestingly, Aeneas hedged its bets,” said Cooley, who praised the model’s accuracy. “In doing so, it exactly reflected the current difference in scholars’ opinions, giving two probable date ranges rather than a single prediction.”
This saga bolsters the idea that AI can be useful for subject matter experts who know what they are doing. (That’s in contrast to when non-experts use AI, when bizarre and sometimes unsettling things seem to happen consistently.)
If users of this program use this tool effectively — and it looks like they can — anybody who’s majorly into ancient history should be very excited; we could soon learn way more about our past.
More on AI: Scientists Are Sneaking Passages Into Research Papers Designed to Trick AI Reviewers