AI Revives Lost Latin Inscriptions: How Aeneas is Reconstructing Ancient Roman Texts!
Every year, around 1,500 ancient Latin inscriptions are uncovered across the former Roman Empire. Whether carved into stone, painted onto walls, or engraved on bronze and marble, these messages offer a rare glimpse into everyday life in ancient Rome. But with time taking its toll, many of these inscriptions have been left fragmented and incomplete, puzzles missing crucial pieces.
Now, thanks to cutting-edge work from Google DeepMind and academic researchers, there’s a new tool helping to fill in the gaps. Meet Aeneas, an artificial intelligence model named after the legendary Trojan hero of Roman myth. Aeneas isn’t just your average AI, it’s been trained to restore lost text, estimate when and where an inscription was made, and even find similar examples, all in a matter of seconds.
The AI was trained on more than 176,000 Latin inscriptions sourced from major epigraphic databases like EDR (Epigraphic Database Roma), Heidelberg, and Clauss-Slaby. That’s over 16 million characters' worth of content, including both written text and imagery. Unlike traditional tools that might just search for matching words, Aeneas recognises patterns across centuries and across the empire.
Photo Credit: Public domain via the Metropolitan Museum of Art.
“What’s so powerful about inscriptions is that they were written by people from all walks of life, soldiers, shopkeepers, parents, not just the elites,” said Yannis Assael, a researcher at DeepMind and co-author of the study. “They’re unfiltered voices from the past.”
Of course, many inscriptions are damaged, and often historians aren’t even sure exactly where or when they were made. As Thea Sommerschield, a University of Nottingham epigrapher who co-developed Aeneas, explains: “Studying these texts is like tackling a massive jigsaw puzzle. You can’t do much with just one piece, even if you can see its shape or colour.”
That’s where Aeneas comes in. It compares new fragments to thousands of known inscriptions, helping suggest plausible reconstructions. When dealing with gaps of up to ten missing characters, the AI achieves an accuracy of around 73%. For larger gaps with no set length, that drops to 58%. Still, impressive for a language nearly 2,000 years old. The tool can also estimate the province where the inscription was created (with 72% accuracy) and date it to within about 13 years.
In one high-profile case, Aeneas was used to analyse the Res Gestae Divi Augusti, an inscription where Emperor Augustus recorded his accomplishments. Scholars still debate when it was written. The AI proposed two possible timeframes: 10–1 BCE and 10–20 CE, both consistent with leading academic theories.
In testing, 23 historians used Aeneas to help interpret inscriptions, and 90% found its suggestions helpful. Still, experts are quick to point out that while Aeneas is an incredibly useful tool, it’s no substitute for human judgement and expertise.
Aeneas builds on DeepMind’s earlier project, Ithaca, which focused on Ancient Greek inscriptions. Like Ithaca, Aeneas is freely accessible. There’s an interactive version available at predictingthepast.com, and both the code and dataset have been made open-source for researchers and enthusiasts alike.
By combining machine learning with historical research, Aeneas helps us hear the voices of the past more clearly and brings us a step closer to understanding the people who once carved their lives into stone.