DeepMind's Aeneas Uses AI to Unlock Ancient Roman Inscriptions

Google DeepMind has introduced Aeneas, the first AI model designed to help historians interpret and restore ancient Roman inscriptions. Published in Nature, the tool represents a major shift in how scholars approach fragmentary texts etched onto monuments, everyday objects, and artifacts across the ancient world.

The Challenge of Fragmentary Texts#

Roman inscriptions are scattered across political graffiti, epitaphs, business records, and even magical texts — but many are weathered, incomplete, or deliberately defaced. Historians traditionally spend months manually searching for "parallels" — similar texts that share wording, formulas, or origins — to understand isolated fragments and establish their context, date, and provenance.

Aeneas accelerates this painstaking work by reasoning across more than 176,000 Latin inscriptions, retrieving contextual matches in seconds.

How It Works#

Aeneas is a multimodal neural network that processes both text and images of inscriptions. The model draws on three major digital epigraphic databases to create what DeepMind calls a "historical fingerprint" of each text — encoding linguistic features, temporal markers, geographical data, and connections to related inscriptions.

The model can:

• Identify parallels across vast collections using deep textual and contextual analysis
• Restore damaged sections with 73% accuracy for gaps up to ten characters, and 58% for unknown-length gaps
• Determine geographical provenance using both text and images, achieving 72% accuracy across 62 ancient Roman provinces
• Date inscriptions within an average of 13 years
• Handle multimodal input — the first model to combine visual and textual analysis for Roman epigraphy

Real-World Application#

When tested on the Res Gestae Divi Augusti — Emperor Augustus' famous first-person account of his achievements — Aeneas produced a probabilistic date distribution that captured both major scholarly hypotheses about when the text was written. Rather than forcing a single answer, it showed two peaks: one around 10–1 BCE and a larger one between 10–20 CE, grounded in subtle linguistic and historical markers.