In the Roman world, writing was etched everywhere—from imperial monuments to everyday objects. Political graffiti, love poems, epitaphs, business transactions, birthday invitations, and magical spells all offer modern historians a rich window into the diversity of daily life. Yet many of these inscriptions are fragmentary, weathered, or deliberately defaced, making restoration, dating, and contextual placement nearly impossible without deeper insights.
Introducing Aeneas—the first artificial intelligence model designed specifically for contextualizing ancient inscriptions. Published in Nature, Aeneas revolutionizes how historians interpret, attribute, and restore fragmentary texts. Traditionally, experts rely on their knowledge and specialized resources to identify “parallels”—texts with similarities in wording, syntax, or provenance. Aeneas accelerates this complex process, reasoning across thousands of Latin inscriptions in seconds and retrieving textual and contextual parallels that empower historians to build upon its findings.
Developed in collaboration with the University of Nottingham, along with researchers from Warwick, Oxford, and Athens University of Economics and Business, Aeneas is part of a broader effort to explore how generative AI can help historians identify and interpret parallels at scale. An interactive version is freely available at predictingthepast.com, and the code and dataset are open-sourced to support further research.
Named after the wandering hero of Graeco-Roman mythology, Aeneas builds upon earlier work like Ithaca, which used AI to restore, date, and place ancient Greek inscriptions. Aeneas goes further by helping historians interpret and contextualize texts, giving meaning to isolated fragments and enabling richer conclusions about ancient history.
Key Capabilities of Aeneas:
- Parallels Search: It scans a vast collection of Latin inscriptions, turning each text into a historical fingerprint to identify deep connections and situate inscriptions within their broader context.
- Multimodal Input Processing: Aeneas is the first model to determine a text’s geographical provenance using both text and visual information, such as images of inscriptions.
- Restoring Gaps of Unknown Length: For the first time, it can restore gaps in texts even when the missing length is unknown, making it highly versatile for heavily damaged materials.
- State-of-the-Art Performance: Aeneas sets new benchmarks in restoring damaged texts and predicting their origins and dates.
Aeneas operates as a multimodal generative neural network, processing both an inscription’s text and image as input. It was trained on the Latin Epigraphic Dataset (LED), which includes over 176,000 Latin inscriptions curated from decades of historical work, including resources like the Epigraphic Database Roma and Epigraphic Database Heidelberg. The model uses a transformer-based decoder to handle textual input, with specialized networks for restoration, dating, and geographical attribution.
In terms of performance, Aeneas groups inscriptions by date far more clearly than general-purpose models trained on Latin. It restores damaged inscriptions with 73% Top-20 accuracy for gaps up to ten characters, and even when the restoration length is unknown, accuracy remains at 58%. Thanks to its use of visual data, it attributes inscriptions to one of 62 ancient Roman provinces with 72% accuracy and dates texts within 13 years of historian-provided ranges.
Aeneas also offers a new lens on historical debates. When tested on the famous Res Gestae Divi Augusti—Emperor Augustus’ first-person account—it produced a detailed distribution of possible dates, capturing prevailing scholarly hypotheses quantitatively. The model based its predictions on subtle linguistic features and historical markers, retrieving parallels from imperial legal texts tied to Augustus’ legacy.
In a collaborative study with 23 historians, Aeneas proved invaluable. It helped identify new parallels and increased confidence in tackling complex epigraphic tasks. Historians noted that Aeneas accelerated their work and expanded the range of relevant parallels. One anonymized historian remarked, “Aeneas’ parallels completely changed my perception of the inscription. It noticed details that made all the difference for restoring and chronologically attributing the text.”
Designed to integrate seamlessly into historians’ workflows, Aeneas combines expert knowledge with machine learning, offering interpretable suggestions that serve as starting points for inquiry. As part of this release, the ancient Greek model Ithaca has been upgraded with Aeneas’ capabilities, including contextualization and restorations of unknown length. A new teaching syllabus has also been co-designed to bridge technical skills with historical thinking, aligning with AI literacy frameworks from the European Commission and UNESCO.
The Aeneas team continues to partner with subject matter experts, using this powerful tool to shed light on our ancient past—with more exciting developments on the horizon.