In a significant leap forward for the digital humanities, Google DeepMind has unveiled Aeneas, a state-of-the-art AI model trained to analyze, restore, contextualize, and connect thousands of ancient Latin inscriptions. Named after the mythic Trojan hero whose journey laid the legendary foundations of Rome, Aeneas marks a new era in the relationship between artificial intelligence and historical research.
A New Approach to Inscriptions
Where previous models like Pythia or Ithaca focused primarily on reconstructing damaged texts, Aeneas introduces a more holistic method. It operates not merely as a restorer but as a contextual interpreter, capable of suggesting parallels, estimating chronology and geography, and identifying recurring patterns across large corpora of epigraphic material.
Trained on a dataset of over 174,000 Latin inscriptions, the model can interpret fragmentary evidence, propose plausible restorations of missing sections, and return semantically similar texts—even when they don’t share explicit vocabulary. This opens new pathways for understanding inscriptions not as isolated finds, but as part of a broader intertextual and historical landscape.
Multimodal Capabilities and Performance
One of Aeneas’ most innovative features is its multimodal architecture. Unlike traditional language models, it can analyze not only textual transcriptions but also images of the inscribed stones themselves. By incorporating visual cues—such as layout, material, and carving style—it achieves high levels of accuracy in dating and regional attribution.
Dating precision: The model can estimate the age of inscriptions within a ±13-year margin, significantly outperforming experts who average ±31 years.
Geographic attribution: It accurately identifies the province of origin for an inscription from among 62 Roman provinces, achieving 72% accuracy using both text and imagery.
Restoration performance: Aeneas demonstrates 73% top-20 accuracy in reconstructing arbitrary-length missing segments—a notable feat in handling fragmentary Latin.
Connecting the Past: The Parallels Function
Perhaps the most transformative function is Aeneas’ ability to suggest relevant parallels. When given a fragmentary or complete inscription, it returns a ranked list of other inscriptions that share structural, thematic, or formulaic features. This capability mimics the interpretive process of expert epigraphers, who draw on years of exposure to formulae, provincial customs, and linguistic patterns to contextualize individual finds.
In practical testing, 23 professional historians collaborated with DeepMind to assess the model’s interpretive suggestions. The use of AI-assisted parallels led to a 44% increase in confidence during scholarly evaluations and was rated helpful in over 90% of cases.
Implications for Scholarship and Methodology
The emergence of Aeneas challenges longstanding boundaries between machine processing and historical interpretation. Rather than supplanting expertise, the model functions as an augmented cognitive tool, offering accelerated comparisons and inferential suggestions that would take weeks or months by conventional means.
Its capacity to identify connections across inscriptions—spanning diverse regions, time periods, and formats—enables new kinds of historiographical synthesis. For example:
A fragment from Germania might now be interpreted alongside an administrative edict from Syria based on shared formulae, shedding light on the bureaucratic commonalities of imperial peripheries.
Dating uncertainties can be narrowed more confidently when contextual parallels and visual style align across the database.
Yet, this power invites caution. The interpretive authority must remain with the scholar, who understands local idiosyncrasies, epigraphic conventions, and historical nuance. AI models offer suggestions, not conclusions—and their results must always be triangulated with material context and expert critique.
Open Access and Future Horizons
DeepMind has made Aeneas accessible via the interactive website predictingthepast.com, with datasets and code openly available. This democratizes access to high-level computational tools, allowing scholars, students, and heritage institutions to integrate AI into both research and teaching environments.
While Aeneas currently focuses on Latin, its architecture lays the groundwork for expansion into other ancient languages—especially Greek, where vast corpora of inscriptions await similar treatment. Extensions to Coptic, Demotic, Aramaic, or even cuneiform would further amplify its significance across Mediterranean and Near Eastern studies.
Aeneas does not merely automate a scholarly task; it reshapes how that task is conceived. It offers a glimpse into a future where humanistic inquiry and machine learning collaborate to illuminate the fragmentary, scattered traces of antiquity with greater coherence and contextual richness than ever before.
It is not a substitute for epigraphy—it is its ally, a digital companion in the ever-evolving pursuit of understanding the written past.
Acknowledgements
The research was co-led by Yannis Assael and Thea Sommerschield
Contributors include: Alison Cooley, Brendan Shillingford, John Pavlopoulos, Priyanka Suresh, Bailey Herms, Jonathan Prag, Alex Mullen and Shakir Mohamed. The Aeneas web interface was developed by Justin Grayston, Benjamin Maynard, and Nicholas Dietrich, and is powered by Google Cloud.
The syllabus was developed by Robbe Wulgaert, Sint-Lievenscollege, Ghent, Belgium.
