AI helps Latin scholars to decipher ancient Roman texts

Around 1,500 Latin inscriptions are discovered each year, offering valuable perspectives on the daily lives of ancient Romans and poses a difficult challenge to historians tasked with interpreting them.

However, a new artificial intelligence tool, partially developed by Google researchers, has now helped Latin scholars put together these puzzles from the past, according to a survey released Wednesday.

Latin inscriptions were commonplace throughout the Roman world, from laying out the decrees of the emperor to graffiti on city streets. A mosaic of one person outside a house in an ancient city of Pompeii even warns, “Beware of the dogs.”

Research co-author Yannis Assael, a researcher at Google's AI Lab Deepmind, said these inscriptions are “invaluable to historians.

“What makes them unique is that they are written by the ancient people themselves across all social classes on any subject. It's not just history written by the elite,” Assael, who co-designed the AI model, told a press conference.

However, these texts are often corrupted for thousands of years.

“We usually don't know where they were written,” Asael said.

Therefore, researchers created generative neural networks. This is an AI tool that can be trained to identify complex relationships between data types.

They named the model Enes, after the Trojan hero and son of the Greek goddess Aphrodite.

They were trained on data on dates, locations and meanings of Latin transcriptions from the empire, spanning 5 million square kilometres over two thousand years.

Thea Sommerschield, an epigrapher at the University of Nottingham, co-designed the AI model, but said, “Studying history through inscriptions is like solving a huge jigsaw puzzle.”

“Even knowing information such as colours and shapes does not allow you to solve puzzles in a single isolated piece,” she explained.

“To solve a puzzle, you need to use that information to find the piece that connects to it.”

Tested at Augustus

This can be a big job.

Latin scholars should compare it to “potentially hundreds of similarities.” This is the study of the journal, “requiring extraordinary learning” tasks and “cumbersome manual searches” through large library and museum collections. Nature I said.

Researchers trained the model with 176,861 inscriptions (value up to 16 million characters). 5% of that included images.

Now, they said that it is possible to estimate the location of the inscriptions among the 62 Roman provinces, provide ten years when produced, and speculate what could be included in the missing section.

To test their model, the team asked Enes to analyze the famous inscription called “Res Gestae Divi Augusti.”

Discussions among historians have been raging about when the text is written accurately.

The text is full of exaggeration, unrelated dates and false geographic references, but researchers said that Eneas could use subtle cues such as archaic spelling to land on two possible dates.

According to DeepMind, more than 20 historians who tried the model provide a useful starting point in 90% of cases.

The best results are when historians use AI models together with their skills as researchers, and instead of relying on one or the other, they use their skills as researchers.

“Since their breakthrough, generative neural networks have appeared to be at odds with educational goals. The fear of relying on AI prevents critical thinking rather than boosting knowledge.”

“Developing Aeneas demonstrates how this technology can meaningfully support the humanities by addressing the specific challenges faced by historians.”

©2025 AFP