A new AI model, Aeneas, can not only estimate origins and dates for ancient Latin texts but also accurately restore lost words, according to a new paper published in Nature.
Some members of the team behind Aeneas previously found success with an AI tool for deciphering Greek texts, which was used on the Herculaneum scrolls buried by the eruption of Mount Vesuvius in 79 AD. While the model shows promise as an addition to the archaeological toolset, critics—and even its creators—caution that studying the past is not yet ready to be fully automated.
Epigraphy
Epigraphy, the study and interpretation of ancient inscriptions, is an essential yet challenging aspect of understanding our ancient past. Traditionally, this work involves the time-consuming process of comparing various inscriptions to identify similarities that might shed light on meaning and origin. Complicating matters further are changes in languages over time and the loss of textual fragments. With a continually growing corpus of known texts, the collective knowledge in the field has become too vast for any one person to master.
As a solution to this overwhelming data landscape, Aeneas was conceived. The team behind the project included researchers from universities in the UK and Greece, working alongside developers at DeepMind, a London-based AI company owned by Google. Drawing from three of the most extensive Latin epigraphical databases, the researchers compiled a massive dataset of 176,861 inscriptions. Five percent of these inscriptions were accompanied by images of the original pieces, which were created between the 7th century BC and the 8th century AD.
Operating Aeneas
The model performs three main tasks, each managed by a dedicated neural network: identifying a text’s place of origin, determining its age, and predicting missing portions of the text. Importantly, Aeneas leaves a transparent trail by listing the relevant examples from known data that it used to generate its results.
“Aeneas can retrieve relevant parallels from across our entire data set instantly,” says Yannis Assael, a research scientist at Google DeepMind and co-author of the new Nature paper.
Aeneas is currently available for public use on Google DeepMind’s history-focused website, Predicting the Past. The present version of the model requires significant human input. The Debrief attempted to clarify why the version available on the website does not accept visual inputs—as mentioned both on the site and in the paper—but received no response from the paper’s authors as of the time of publication.
“While the PR makes it seem like you feed Aeneas a photograph of a broken tablet and it will recreate the rest, the actual work it does is much less impressive,” said Jason Colavito, an author, editor, and independent historical researcher who was not involved with the Nature study, in an email to The Debrief where he discussed his initial tests of the tool.
“Users are asked to manually transcribe and enter the partial inscription as typed text,” said Colavito, adding that they are “asked to estimate how many characters are missing from each line, so already the user is doing a good chunk of the heavy lifting before the A.I. even gets started.”
Internal Testing Results
To test Aeneas, researchers compared its performance to that of 23 expert human epigraphers working without AI assistance, as well as to experts using the model as a research aid. The results showed that human experts alone produced dating estimates with an average deviation of 31 years, while Aeneas alone narrowed the margin to 13 years. When experts used Aeneas’s suggested reference data, their accuracy improved to within 14 years. For more complex tasks such as identifying geographical origin and restoring missing text, the highest accuracy came from humans using Aeneas as a tool, rather than working independently.
One test involved the Roman imperial text Res gestae divi Augusti, documenting the life of Emperor Augustus. Notably, the internal dates in the document did not confuse the AI, which produced results comparable to those of human experts in terms of dating. Aeneas also successfully identified variant spellings and linguistic cues often used by human epigraphers to determine a text’s geographic and chronological origin.
Further testing on a Latin-inscribed altar also delivered promising results. Aeneas correctly identified another altar from the same region as being related, despite receiving no external information indicating a connection between the two.
University of Sydney Latin scholar Anne Rogerson said in a comment provided to Nature that Aeneas represents an improvement over previous AI models, which have been known to hallucinate answers. “It’s giving a hypothesis based on the evidence base that it’s working from, so it’s a rational guess rather than a wild stab in the dark.”
Issues with Aeneas
Despite its capabilities, the Aeneas team offers an important caveat. Like all AI systems, Aeneas doesn’t generate original insights—it compares existing data. While the dataset used to train Aeneas was large by archaeological standards, it remains significantly smaller than the training data for broader models, such as ChatGPT. This limitation means that unconventional inscriptions, particularly those from underrepresented time periods or regions, may pose challenges for the model.
“The formulaic nature of many Roman inscriptions is a particularly good fit for AI-facilitated tools,” Dr. Charles Kuper of the University of Tennessee told The Debrief, agreeing with the team’s view that non-standard texts may remain problematic.
Colavito told The Debrief that he tested Aeneas by inputting a partially redacted Roman inscription. While the AI accurately identified the style and date, it struggled with restoring proper names and numbers but performed well with common words, similar to how a human might fill in missing text.
“I tested Aeneas by taking a complete inscription found in Germany and dated to the first century CE and giving Aeneas about two-thirds of it, with some words and characters missing, to see how it would restore it,” Colavito explained. “Aeneas correctly identified the style and the approximate date of the inscription, but the restoration was not even close on proper names or numbers used in the inscription, though for more common words, it correctly filled in missing letters, as a human scholar would undoubtedly also do.”
Despite its flaws, Colavito acknowledges the model’s value as a supplemental research tool.
“It can be a useful starting point to suggest likely avenues for classification and research, which seems to be Google’s intention for how it should be used,” he said. “They even say on the Aeneas website that historians considered Aeneas’s results to be ‘useful research starting points,’ but they don’t say that the results are definitive or even accurate.”
The paper “Contextualizing Ancient Texts with Generative Neural Networks” appeared on July 23, 2025, in Nature.
Ryan Whalen covers science and technology for The Debrief. He holds an MA in History and a Master of Library and Information Science with a certificate in Data Science. He can be contacted at ryan@thedebrief.org, and follow him on Twitter @mdntwvlf.