TLDR: Google DeepMind, in collaboration with the University of Nottingham, has unveiled ‘Aeneas,’ a new generative neural network capable of predicting missing words, dating, and geographically locating ancient Latin inscriptions. Published in Nature, this AI tool significantly enhances the efficiency and accuracy of historical research, offering a powerful aid to epigraphers and classicists.
A groundbreaking development in the field of digital humanities has emerged with the introduction of ‘Aeneas,’ a novel generative neural network designed to contextualize ancient texts. Developed by Google DeepMind in collaboration with researchers, including those from the University of Nottingham, this AI model, detailed in a paper published in the prestigious journal Nature on July 23, 2025, promises to transform the study of ancient Latin inscriptions.
Each year, approximately 1,500 new Latin inscriptions are discovered, offering invaluable insights into the language, culture, and daily life of the Roman Empire. However, these texts are often fragmented, with missing letters, words, or even entire sections due to the ravages of time and intentional destruction. Traditionally, the restoration, dating, and geographical attribution of these inscriptions have been a laborious and highly specialized manual process, requiring experts to compare target inscriptions with hundreds of similar ‘parallels’ from vast historical datasets. This method is not only time-consuming but also limits the ability of scholars to identify large-scale historical connections.
Aeneas addresses these challenges by employing a multi-modal approach, analyzing both the visual characteristics of the inscribed object and the text itself. The model is capable of predicting absent words, determining the approximate date of creation, and pinpointing the geographical origin of the inscriptions. According to the research, Aeneas provides useful answers in 90% of cases and has been shown to optimize tasks by an impressive 44%. Furthermore, it can provide datings within an average error margin of just 13 years, a remarkable feat attributed to its ability to locate texts in specific times and places based on its learned dataset.
Dr. Thea Sommerschield, an epigrapher at the University of Nottingham and a co-developer of Aeneas, expressed her astonishment at the model’s capabilities, particularly its ability to discern subtle linguistic and historical markers. Yannis Assael, a key researcher involved in the project, stated, ‘The ability of Aeneas to help scholars contextualize and restore their materials has great possibilities to generate new lines of research in our field.’ Charlotte Tupman from the University of Exeter further noted the tool’s immense potential, extending beyond ancient history to later inscriptions and other languages.
In a significant test, Aeneas was tasked with analyzing ‘Res Gestae Divi Augusti,’ a lengthy inscription detailing the accomplishments of Rome’s first emperor, Augustus. The AI did not provide a single answer for its creation date but instead showed a probability distribution with two peaks, aligning with two prominent academic hypotheses. This demonstrated Aeneas’s capacity to grasp nuanced linguistic and historical markers, such as archaic spellings or specific official titles, much like a human expert.
While Aeneas represents a significant leap forward, researchers and scholars acknowledge its limitations. Kathleen Coleman, a professor of classics at Harvard University, highlighted that Aeneas ‘does not infer the meaning of the text,’ emphasizing that interpretation remains a human endeavor. The long-term utility and the model’s performance on more obscure fragments, as opposed to well-studied samples, are still subjects for further investigation. A fundamental challenge also lies in the inherent biases within the training dataset, as surviving inscriptions do not fully reflect the entirety of the ancient Roman world. Researchers stress that Aeneas is a powerful tool to be used critically by historians, with the ultimate responsibility for interpretation resting with human experts.
Also Read:
- DeepMind’s AlphaEvolve: An AI System Revolutionizing Algorithmic Discovery and Optimization
- AI Breakthrough: Google and OpenAI Models Achieve Gold Medal Performance at International Math Olympiad
Future research aims to develop better methods for representing and evaluating broad date ranges, enhance multi-modal capabilities with larger and more standardized datasets, and expand the scope beyond Latin inscriptions, fostering deeper interdisciplinary collaboration between humanities and natural sciences.
