Newsletters
A new AI-powered tool is transforming the way historians interpret damaged Latin inscriptions from the ancient Roman world.
Developed by researchers from Google DeepMind and leading universities in the U.K. and Europe, the model—named Aeneas—can restore missing texts, date inscriptions, and determine their geographical origins with unprecedented precision.
Latin inscriptions, carved into stone and metal across the vast expanse of the Roman Empire, provide critical insight into its language, culture and political order.
However, many of these inscriptions have been damaged over time—entire sections are missing, eroded or illegible.
Historically, scholars have had to rely on painstaking manual methods to reconstruct these gaps, often limited by access to source materials and expertise in specific regional or chronological areas.
With Aeneas, researchers can now draw upon AI-generated parallels to support their interpretations, offering a scalable solution to a longstanding academic bottleneck.
Unlike earlier models, Aeneas doesn't just match inscriptions by keywords or known patterns.
Instead, it uses a combination of text and image inputs to retrieve meaningful parallels across historical periods and regions, even when the missing text length is unknown.
By training on over 176,000 Latin inscriptions collected from public databases across Europe, Aeneas constructs a rich “embedding space” that captures the stylistic, linguistic, and contextual similarities among ancient texts.
Its outputs include ranked lists of similar inscriptions, restoration hypotheses, and metadata on date and provenance.
The name Aeneas references the mythical Trojan hero who wandered across the Mediterranean—just as the model “journeys” across space and time to find meaningful connections between fragmented inscriptions.
To assess its utility, Aeneas was tested in a large-scale study involving 23 expert historians. In controlled experiments, participants were asked to restore, date, and geographically attribute 60 inscriptions under three scenarios: working alone, working with parallels suggested by Aeneas, and working with both parallels and the model’s predictive outputs.
Results show that historians who used Aeneas’ assistance significantly outperformed those who worked unaided.
In restoration tasks, character error rates dropped from 39% to 21% when using both parallels and predictions.
Geographical accuracy improved from 27% to 68%, and average error in dating inscriptions fell from 31.3 years to 14.1 years—nearly matching the AI's standalone performance of 12.8 years.
Participants also reported higher confidence levels and found the suggested parallels valuable starting points in 90% of cases.
To illustrate its capabilities, researchers applied Aeneas to the Res Gestae Divi Augusti, the autobiographical inscription by Emperor Augustus.
Carved onto the Temple of Rome and Augustus in Ankara (ancient Ancyra), it is one of the most famous political texts from antiquity.
Aeneas identified key chronological markers in the text, such as references to titles and monuments associated with specific dates.
It also retrieved parallels from other official Roman inscriptions, including senatorial decrees that shared similar stylistic and ideological traits. These AI-generated insights aligned closely with established scholarly interpretations, showcasing the model’s reliability.
The team emphasizes that Aeneas is not a replacement for human expertise but a collaborative tool that enhances traditional research methods.
Its performance depends on extensive training data, but its architecture can be adapted to other ancient languages and scripts.
Currently available through an online interface, Aeneas may soon help researchers across the globe engage with epigraphy, papyrology and even coinage studies in new ways.
2 min read
