Conversing with antiquity: Agentic AI partner for expanding historical research

Conversing with antiquity

An agentic AI partner for expanding historical research

Thea Sommerschield, Durham University

Zoi Tsangalidou and Yannis Assael, Google DeepMind

Ancient inscriptions offer a direct window into the human past. As invaluable historical sources, they preserve everything from imperial decrees to everyday transactions of ordinary citizens. However, many inscriptions have been damaged to the point of illegibility, their date and place of writing steeped in uncertainty. Reassembling these broken narratives is one of humanity’s greatest challenges, requiring expert historians to solve complex, localized puzzles of text, time and space.

For nearly a decade, we have partnered closely with epigraphers to pioneer state of the art AI tools for historical research. The milestones in this journey include Ithaca (2022) and Aeneas (2025), our generative models for restoring, dating and placing ancient Greek and Latin inscriptions. To place these capabilities directly into the hands of researchers, we previously built an interactive online platform at predictingthepast.com and fully open-sourced our underlying models.

Community collaboration has highlighted three main challenges for AI-assisted historical analysis. First, preserving explainable, flexible interpretations requires tailored visualizations for individual inscriptions. Second, advanced multi-text analysis must move beyond basic comparisons without requiring specialized coding. Finally, large language models must be firmly grounded in evidence and expertise to remain reliable.

To overcome these barriers, the Predicting the Past Skill for Google Antigravity shifts these complex computational workflows into natural language. By grounding Gemini directly in the specialized outputs of Aeneas and Ithaca, we have created an interactive partner that allows historians to attribute, restore, and analyze ancient texts as naturally as having a conversation with a colleague.

To demonstrate the practical power of this collaborative approach, we worked closely with Dr. Thea Sommerschield, a historian and epigrapher at Durham University, who has co-led all projects in this domain. Together we put the system to work across three distinct, real-world case studies that span the Greco-Roman worlds, show how researchers can now perform large-scale, interactive, and visually rich epigraphic analyses.

The ring thief of Aquae Sulis (Tab.Sulis 97)

To examine how the system handles formally and stylistically distinctive language, and how it supports explainability, we deployed the model to attribute and analyze a Latin curse tablet (defixio) from Roman Britain, recovered from a votive deposit at the hot-spring sanctuary of Minerva at Bath (Aquae Sulis). The inscription was written by a woman named Basilia, cursing whoever had stolen her silver ring. Hundreds of similar curse tablets have been found at the site, making Bath one of the richest sources of this distinctive epigraphic genre.

Curse tablet with complaint about the theft of Vilbia. Photograph by Mike Peel.

In this case, Aeneas places the inscription within the chronological and geographical ranges proposed by historians, while providing a transparent account of how that conclusion was reached. More interestingly, the explanation it produces begins to resemble a piece of epigraphic commentary in its own right: an interpretation of the textual features that underpin historical attribution.

Mapping the cult of the Aufaniae (CIL XIII, 6665)

Reconstructing damaged text and analyzing regional patterns across a wider corpus can be demonstrated through the example of a votive altar from Mainz (Mogontiacum), dedicated in 211 CE by the provincial official (beneficiarius consularis) Lucius Maiorius Cogitatus. The inscription honours a group of Germanic mother-goddesses, the Aufaniae. Many similar dedications are known from the Rhine and Danube provinces, often set up by Roman soldiers and administrators. Although these inscriptions are highly formulaic, small differences in their wording can reveal regional traditions and patterns of mobility across the Empire.

Heidelberger Akademie der Wissenschaften/Epigraphische Datenbank Heidelberg: © AG-OS Fremdfoto

In this example, the Skill moves beyond the analysis of a single inscription, while remaining grounded in evidence that historians can inspect and verify. It identifies patterns across an entire corpus and traces how religious practices travelled through the movement of people across the Roman Empire.

Who came to Dodona?

Processing bulk collections of fragmentary inscriptions and mapping high-dimensional semantic relationships using the skill is illustrated through the lead oracular tablets of Dodona in northwest Greece. Like Delphi, Dodona attracted visitors from across the Greek world seeking divine guidance on matters ranging from business and travel to family affairs and religious obligations.

Thousands of lead tablets survive from the sanctuary, preserving the questions that ordinary people brought to the oracle. Many are extremely fragmentary, but together they provide an unparalleled record of everyday concerns in the ancient Mediterranean.

Lead plaque asking questions of an oracle at Dodona. Zde, CC BY-SA 4.0, via Wikimedia Commons.

In this example, the Skill moves beyond attributing individual inscriptions. It reconstructs the wider community of people who visited Dodona, allowing historians to explore the sanctuary not only as a collection of texts, but as a network of connected individuals moving through the ancient Mediterranean.

The Predicting the Past skill can help historians use advanced AI models through natural language and, by unblocking some of the complex bottlenecks, allows epigraphers to analyze patterns and produce visualizations in a matter of minutes. By bridging the specialized outputs of Aeneas and Ithaca with Gemini's interactive reasoning, this collaborative AI partner brings together multiple other tools that can now be directly integrated into historians’ workflows and help expand the boundaries of their research.

Try the skill

Acknowledgements

We would like to acknowledge the contributions of the following team members: Saz Basu, Doug Fritz, Tanya Marwah, Sebastian Nowozin, Michael O'Neill, Filip Radlinski, Sukhdeep Singh, Artem Sobolev, Ashok Thillaisundaram, Yunhan Xu.

The Ithaca and Aeneas projects rely on the availability of high-quality datasets of ancient Greek and Latin inscriptions. These datasets were built through centuries of scholarly collection and decades of digital editorial work.

The analysis of the oracular tablets from Dodona was made possible by Dodona Online, a project providing critical reeditions of the oracular tablets as I.Dodone Online, available at dodonaonline.com. We would like to thank the Dodona Online project team for their invaluable work in creating and maintaining this resource.

Ithaca draws on a 2019 version of the Searchable Greek Inscriptions database made available under "Fair Use" license by the Packard Humanities Institute, generously supported by David Packard: inscriptions.packhum.org.

Aeneas was trained on data from: the Epigraphic Database Roma (EDR, available at edr-edr.it), made available pursuant to a Creative Commons Attribution 4.0 International License (CC-BY) on Zenodo; the Epigraphic Database Heidelberg (EDH, available at edh.ub.uni-heidelberg.de), made available pursuant to a Creative Commons Attribution-ShareAlike 4.0 International License (CC-BY-SA) on Zenodo; the ETL repository for the Epigraphic Database Clauss Slaby (EDCS_ETL, available at manfredclauss.de and github.com/sdam-au/EDCS_ETL), made available pursuant to a Creative Commons Attribution 4.0 International License (CC-BY) on Zenodo.

These resources bring together a large proportion of published inscriptions in a searchable digital format.

Aeneas transforms how historians connect the past

July 2025Research

Learn more

Predicting the past with Ithaca

March 2022Research

Learn more