The Romans created a remarkably organized and well-structured society. A big part of that was writing; they wrote often, and they wrote descriptively. They had records of Roman citizens, trade agreements, legal trials, laws, and many, many other things. And they carved these into stone walls, bronze plaques, urns, and even lead curse tablets. These texts are some of our richest sources of the details of daily life in antiquity. But they rarely survive intact.
For historians, piecing things together is painstaking work. They have to find context in fragments, link them with other existing records, and often rely on encyclopedic knowledge and laborious manual searches. Now, a new AI tool is helping to turn back the clock.
A team from Google DeepMind and several universities introduces Aeneas, a generative (and free) AI model designed to help historians contextualize ancient Latin inscriptions. The AI’s core function is to identify “parallels” — other inscriptions with similar wording, function, or cultural settings.
An archaeological detective
The “big idea” behind Aeneas is a process called contextualization. Think of an ancient inscription as a single puzzle piece. With just this, it’s almost impossible to discern the big picture. To truly understand it, you need to find other pieces it connects to. Historians do this by searching for the above-mentioned parallels.
Aeneas was trained on the Latin Epigraphic Dataset (LED), a massive corpus of over 176,000 inscriptions compiled from three major databases. The model converts each inscription into a sort of digital fingerprint that captures not just its text but also its historical and linguistic patterns. By comparing these fingerprints, Aeneas can instantly retrieve a ranked list of the most relevant parallels to help a historian ground their research.
Simply put, the AI takes text (and in some instances, images) and builds a list of related inscriptions. It doesn’t just search for similar words, it also identifies and links inscriptions through linguistic similarities and other connections.
To test its real-world value, the researchers conducted the largest collaborative study between ancient historians and AI to date, involving 23 experts. The results demonstrated a powerful synergy. When working alone, historians achieved a 39% character error rate in restoring texts. With Aeneas’s predictions, that error rate dropped to 21%, outperforming the model working by itself. The tool boosted historians’ confidence by 44% and was deemed a useful starting point for research in 90% of cases.
For instance, one unnamed expert noted:
“The parallels retrieved by Aeneas completely changed my perception of the (evaluated) inscription. I did not notice details that made all the difference in both restoring and chronologically attributing the text.” Similarly, another reported: “The help of parallel inscriptions is great for understanding the type of inscription of fellow soldiers setting up inscriptions, whereas my own search became more narrow zoning in on a set of inscriptions.”
A historian’s dream assistant
The AI was named after Aeneas, a prominent figure in both Roman and Greek mythology. He was a Trojan hero who eventually went on to become the legendary founder of Rome.
Aeneas builds on DeepMind’s earlier Greek model, Ithaca, which focused on Greek manuscripts. It could also be adapted for other ancient languages like Hebrew, Coptic, Sanskrit, or Babylonian. It could help reconstruct lost histories, address long-standing scholarly assumptions, and shine light on marginalized voices etched into stone but nearly erased by time. It’s also most helpful for researchers working in understaffed institutions or who don’t have an extensive knowledge of Roman inscriptions.
Developed by DeepMind in partnership with historians from the University of Nottingham, Oxford, and Warwick, Aeneas is multimodal, meaning it can analyze both text and images of inscriptions to improve the accuracy of its predictions.
Researchers also tested it on Res Gestae Divi Augusti, the self-authored obituary of Emperor Augustus, one of the most hotly debated Roman inscriptions. For centuries, historians have debated exactly when it was written. Without prior knowledge, Aeneas analyzed the full text and offered a distribution of possible dates that neatly captured both sides of the debate.
This goes to show both the advantages and the limitations: it offers contextualization and useful information, but you still need the human experts to draw conclusions.
The researchers have open-sourced the model, its code, and the dataset freely available online, opening the door for new discoveries about the Roman world and beyond.
“The Aeneas team is continuing to partner with diverse subject matter experts, using Aeneas to help shed light to our ancient past — with more to come,” writes the DeepMind team.
The study was published in Nature.
