DeepMind's Aeneas helps historians decode ancient Rome

Deciphering broken stones from antiquity has always been the ultimate test of human memory. Epigraphists often spend months digging through library catalogs to find a single inscription with similar phrasing, just to guess a missing word.

Now Google DeepMind wants to turn that search into a matter of seconds. Co-developed with the University of Nottingham, Oxford, Warwick, and Athens University of Economics and Business, DeepMind just published a paper in Nature introducing Aeneas, a multimodal generative neural network built specifically to contextualize Latin inscriptions.

Fingerprinting ancient stone

Roughly 1,500 new Latin inscriptions are unearthed every year, but most are battered, deliberately defaced, or heavily abbreviated. Traditional digital tools basically rely on literal string matching, which collapses the moment letters are missing.

Aeneas takes a completely different path. The researchers unified three massive epigraphic databases into a clean corpus dubbed the Latin Epigraphic Dataset (LED), spanning 176,861 inscriptions and 16 million characters from the 7th century BCE to the 8th century CE.

Two big breakthroughs stand out. First, Aeneas can restore text fragments of arbitrary length. Previous machine learning models required scholars to know exactly how many characters were missing; Aeneas can tackle gaps where the missing length is totally unknown.

Second, it is multimodal. Instead of only analyzing transcribed Latin, Aeneas can look at an image of the stone alongside the text to predict its geographical origin. At the same time, it turns each inscription into an embedding—a kind of historical fingerprint—to surface contextual and stylistic parallels across the Roman Empire in seconds.

When historians team up with AI

In evaluations, historians reported that the parallel texts retrieved by Aeneas served as useful research starting points in 90% of cases, boosting scholarly confidence in interpretations by 44%.

For dating, Aeneas pinned down inscriptions to within 13 years of their ground-truth historical ranges. Crucially, the combined team of human historians working with Aeneas outperformed both solo human specialists and standalone AI models in both text restoration and geographical attribution.

To ensure wide access, DeepMind open-sourced the code and the LED dataset, while deploying an interactive tool for researchers and students at predictingthepast.com.

Why it matters

For historians, epigraphists, and museum curators, Aeneas is a genuine workflow accelerator. It takes the grueling legwork of cross-referencing fragmentary Latin formulas across an empire spanning five million square kilometers and condenses it into an instant list of plausible leads. Instead of replacing the expert, it gives them high-probability starting points, especially on broken stones where letter counts were previously pure guesswork.

From an industry standpoint, we think DeepMind made a very smart design choice here. Rather than building a black-box text generator prone to hallucinating Roman history, they anchored the model around parallel retrieval. By showing scholars existing, verified inscriptions that share the same DNA, the tool provides transparent historical grounding rather than unverifiable text.

The main vulnerability looks to be data distribution. Only 5% of the inscriptions in the training corpus currently include matching imagery, meaning the vision-based geographical predictions still rely on a relatively narrow foundation. Furthermore, the surviving epigraphic record is inherently biased toward wealthier provinces. Still, as a template for how frontier AI can collaborate with specialized humanities scholarship, Aeneas sets a remarkably solid benchmark.