Latin inscriptions from ancient times provide insights into the proclamations of the Roman emperor the voices of the enslaved for those who understand them. Today, advancements in artificial intelligence assist historians in piecing together often incomplete texts, even allowing for predictions about the origins and contexts of specific inscriptions within the Roman Empire.

“Analyzing history through inscriptions resembles solving an enormous puzzle. There are tens of thousands of them,” stated Thea Sommerschield during a media briefing at the University of Nottingham, UK. “This is everything that has endured through the centuries.”

The AI tools developed by Sommerschield and her team can identify missing characters in Latin inscriptions and indicate that they relate to the same individual or location as similarly styled inscriptions. They named the tool Aeneas, after the legendary hero reputed to have fled the fall of Troy, eventually becoming an ancestor to the Romans.

“Aeneas can help restore unknown lengths of text that are missing,” remarked Giannis Assael, co-leader of Aeneas’s development at Google DeepMind. “This versatility is beneficial for historians, particularly with heavily damaged materials.”

The team trained Aeneas using the most extensive dataset of ancient Latin texts, comprising over 176,000 inscriptions and nearly 9,000 accompanying images. This training enables Aeneas to propose missing phrases. Furthermore, when tested against a set of known inscriptions, researchers found that Aeneas could estimate inscription dates within a 13-year range and achieved a 72% accuracy rate in discerning the origins of different regions in Rome.

“Inscriptions are crucial for understanding the lives and experiences of those in the Roman Empire,” noted Charlotte Tapman, a researcher at the University of Exeter, UK, who was not directly involved in this study. “They encompass a variety of topics including law, trade, military affairs, political life, religion, death, and personal matters.”

Such AI technologies “are poised to be implemented for studying inscriptions from other eras and adapted for various languages,” according to Tupman.

In tests involving intentionally damaged inscriptions, Aeneas achieved a 73% accuracy in filling gaps of up to 10 Latin characters. However, when the total length of missing text was unknown, accuracy dropped to 58%. Nonetheless, the AI’s logic for its suggestions enables researchers to evaluate the legitimacy of its findings.

When nearly 20 historians evaluated AI-assisted restorations on intentionally damaged inscriptions, those using AI outperformed both the individual historians and the AI itself. Furthermore, historians reported that the comparative inscriptions highlighted by Aeneas served as valuable starting points for 90% of their research.

“Utilizing this technology accelerates work related to inscriptions, especially when seeking to build broader analyses of local or empire-wide patterns and epigraphic traditions,” stated Elizabeth Meyer from the University of Virginia. “However, it remains essential for human scholars to examine the results and ensure their plausibility within the given historical context.”

“Employing a generic AI model to assist in ancient historical tasks often yields unsatisfactory results,” commented Chiara Senati from the University of Vienna, Austria. “Thus, the creation of tools specifically catered to Latin epigraphy research is greatly valued.”

The ultimate vision is for historians to be able to “redirect their energies towards museums and archaeological sites,” Sommerschield shared during a press event.