Approximately 1,500 ancient Latin inscriptions are discovered each year across the former Roman world—on walls, mosaics, pottery shards, and gravestones. Many of these inscriptions are incomplete, cracked, or damaged, posing a significant challenge to historians. Now, thanks to a new artificial intelligence system called Aeneas, it is possible to reassemble the historical puzzle and reconstruct lost texts from antiquity.
Yannis Assael, an AI expert at Google DeepMind, explains that Latin inscriptions hold immense value for understanding Roman history, society, and language: "They are direct testimony from the past. These are not just records from the elite—they represent writings from people of various classes and topics, making them a rich and unique social document."
However, working with these texts is highly complex. Many are fragmented, written in non-standard Latin, and lack clear dating or geographic context. To tackle this challenge, researchers developed the Aeneas model—a generative multi-modal neural network published in the journal Nature. The model uses artificial intelligence to identify connections between different inscriptions and reconstruct their historical context.
Thea Sommerschield, a researcher at the University of Nottingham and an expert in epigraphy (the study of ancient inscriptions), who co-developed the model, likens the work with ancient inscriptions to solving a puzzle: "You can't understand everything from a single piece, even if it’s clear. You need to find the other pieces that connect to it in order to see the full picture."
The model is based on three databases that together contain 176,861 Latin inscriptions—about 16 million characters. Around 5% of these include photographs, which enabled the model to factor in additional visual elements such as material, color, and visual symbols. This capability makes Aeneas the first of its kind to rely on both text and the physical features of inscriptions.
The model's goal is to identify similarities among different inscriptions—names, words, phrases, or locations—and use them to determine where and when a given inscription was written. Among other abilities, Aeneas can estimate the origin of an inscription from among 62 Roman provinces, date it by decade, and even suggest reconstructions for missing parts of the text. It also indicates which textual or visual elements influenced its prediction.
Researchers emphasize that in the past, this comparative work was done manually—each inscription had to be examined against hundreds of others, relying on vast knowledge, access to rare libraries, and painstaking effort. Now, the process is significantly faster, allowing historians to focus on analysis rather than technical reconstruction.
As a demonstration, Aeneas was asked to analyze the historical text Res Gestae Divi Augusti—a propaganda document in which Roman Emperor Augustus described his life and achievements. The document, which was copied in many regions throughout the Roman Empire, presents a complex challenge: It includes misleading descriptions, incorrect references, and inconsistent dates. Nevertheless, Aeneas was able to identify clues pointing to two possible dates for its composition—exactly as noted by contemporary scholars.
According to DeepMind data, about 90% of historians who reviewed the system’s predictions reported that the model helped them as a starting point for independent research. In addition, researchers using the model achieved greater accuracy than humans alone or other non-multi-modal AI systems.
Professor Rob Voulgaris, an AI expert at St. Levin College in Ghent, Belgium, and a co-author of the study, believes the Aeneas project challenges outdated perceptions of AI's role in the humanities: “AI systems were once seen as a threat to critical thinking, out of fear they would replace it. But Aeneas proves that advanced technology can support the study of history—not replace researchers, but strengthen their ability to understand the past.”