A broken Roman inscription can be matched with other ancient texts even when the connection goes beyond identical words. Aeneas, an AI model described in Nature in July 2025, searches for those parallels and suggests restorations, dates and locations.
The practical problem is context. A few surviving letters can fit several phrases. Finding another inscription with a related formula, purpose or historical setting can narrow the possibilities.
The researchers assembled a dataset containing more than 176,000 Latin inscriptions. Aeneas turns their text into numerical representations that make related inscriptions easier to retrieve. It can also use photographs when estimating geographical origin.
One unusual feature handles gaps whose original length is unknown. Many restoration tasks start by specifying how many characters are missing. A broken stone does not always preserve that information.
The paper reports that the strongest restoration and geographical results came from historians working with Aeneas. Those combinations outperformed either historians or the model working alone in the study.
Dating needs careful wording. The model's reported distance from reference date ranges was about 13 years. That does not mean every inscription has been assigned its true year within a 13-year margin.
The team released code and an interactive research interface. Suggested words remain hypotheses to examine alongside the object and its archaeological record. The model supplies comparisons; the stone remains evidence.
