Imagine holding a broken stone slab, its Latin text partly eroded by centuries of weather and time. Just enough letters remain to hint at meaning, but not enough to be certain. For decades, archaeologists and historians have faced this frustrating puzzle—how to read what was never fully preserved. Now, thanks to a collaboration between historians and Google DeepMind, a new AI tool called Aeneas is making it possible to complete these ancient conversations.
Named after the legendary Trojan hero of Roman lore, Aeneas isn’t a mythic warrior—it’s a machine learning system trained on over 170,000 ancient Latin inscriptions. Its job? Help scholars restore damaged texts by predicting missing words, even when the length of the missing part is unknown. The tool uses patterns learned from a vast corpus of classical writing and matches fragments to potential meanings with uncanny accuracy. In test scenarios involving simulated damage, Aeneas not only guessed the missing words but did so with a precision that surprised its human collaborators.
Where previous reconstruction efforts often leaned heavily on scholarly intuition and sparse analogs, Aeneas works differently. It doesn’t just guess; it compares, calculates, and connects. One of its most useful features is linking an incomplete inscription to others with similar language, names, or places—offering researchers valuable new leads. Historians reported that these connections were helpful 90% of the time when trying to understand the broader context of a damaged text.
Aeneas can also estimate the origin of an inscription in both time and geography. It can date many texts to within a decade and determine their regional origins with impressive reliability—around 75% accuracy in trials. For researchers dealing with the fragmented realities of ancient Roman life, that’s a huge leap forward. What once took years of cross-referencing and debate can now be done in a matter of hours, accelerating the pace of discovery and deepening our understanding of everyday Roman experiences—from legal edicts and military honors to personal dedications.
The implications go beyond Roman studies. While Aeneas was trained specifically on Latin, its architecture suggests similar models could be built for other languages lost to time. The developers have made the tool free and available to the academic community, opening the door to broader applications.
Of course, human oversight remains essential. Experts emphasize that AI should suggest, not decide. Every restored word must still be tested against what is known about the artifact’s context, location, and historical background. Still, the partnership between machine learning and human scholarship is proving powerful.
The full story, and more details about the project, can be found here.
From the sunbaked stones of Pompeii to forgotten rural outposts, thousands of Roman inscriptions have been whispering incomplete stories for centuries. Now, with AI lending a hand, we’re starting to hear them again—one phrase, one name, one sentiment at a time.