How Does AI Track Ancient Roman Thieves? A Deep Dive into Antigravity's Historical Reasoning

When AI Becomes a Historical Detective

Google's Antigravity platform recently showcased a skill called "Predicting the Past." Unlike conventional AI applications focused on forecasting future trends, this skill turns the tables — directing machine learning's reasoning capabilities toward historical mysteries that have remained unsolved for millennia.

Antigravity is an experimental AI application platform incubated within Google. Its core design philosophy packages complex AI capabilities into modular, reusable "Skill" units. This architectural approach closely mirrors the "AI Agent" paradigm that has gained traction in recent years: rather than relying on a single monolithic model, the skill architecture allows domain-specific reasoning pipelines to be pre-engineered, enabling researchers like archaeologists or classical scholars — who may have little to no machine learning background — to invoke powerful AI reasoning tools as easily as running a search query. This reflects a broader industry trend: the evolution of AI toolchains from "developer-facing" to "domain expert-facing."

Using this skill, researchers accomplished three remarkably dramatic tasks: tracking an ancient Roman ring thief, mapping the spread of secret cults (mystery religions) across Europe, and reconstructing the social networks of visitors to Greek oracles. These cases sound like the work of a historian crossed with a detective — and AI plays the role of the central reasoning engine.

Tracking the Theft of a Roman Ring

Ancient Romans had a distinctive custom: when something was stolen, the victim would write a "curse tablet" (tabella defixionis), denouncing the thief to the gods and calling for divine punishment. These inscriptions, pressed into lead sheets, have survived as firsthand evidence across two thousand years.

Curse tablets (Latin: tabella defixionis; Greek: κατάδεσμος) represent a unique religious-legal practice in the Greco-Roman world. Victims would inscribe or press their accusations onto thin sheets of lead or tin, then cast them into temple wells, bury them underground, or nail them to sacred sites. Over twelve thousand have been excavated across Europe, with the roughly 130 found at the Roman baths in Bath, England, being particularly well-known. These artifacts document everyday events — thefts, curses, marital disputes — and provide extensive records of ordinary Roman citizens' names, occupations, and social relationships: people who rarely appear in formal historical texts. For AI, curse tablets are valuable because of their highly structured narrative pattern (plaintiff–defendant–deity–request), which facilitates information extraction. At the same time, they present significant challenges: Latin abbreviations, regional dialects, and physical damage all complicate analysis.

Using Antigravity's historical reasoning capabilities, researchers were able to cross-reference dispersed curse tablet texts, excavation sites, and name records to reconstruct the full story of a Roman ring theft from fragmented evidence. At its core, this is a large-scale information correlation and pattern recognition task — the AI must find hidden clues pointing to the same event or the same individual within noisy, semantically ambiguous ancient texts.

Done manually by human scholars, this kind of work can take years of archival comparison. The value of AI lies in its ability to generate cross-document associative hypotheses in a fraction of the time, leaving researchers to focus on verification.

Mapping the Spread of Ancient Cults Across Europe

The second case takes a much broader view: charting the geographic distribution of an ancient mystery cult across the entirety of Europe.

The spread of ancient religious cults was rarely systematically documented. Archaeologists must piece together their diffusion paths from scattered inscriptions, temple ruins, and ritual artifacts. This is fundamentally a problem of reconstructing a spatial-temporal network:

• Geographic dimension: The distribution of different sites on a map hints at the routes along which a belief system traveled.
• Temporal dimension: Differences in the dating of remains reveal the sequence of diffusion.
• Textual dimension: Divine names and ritual descriptions in inscriptions help determine whether remains belong to the same cultic tradition.

AI's role here is to unify multi-modal, multi-source evidence within a single analytical framework, ultimately producing a visualized diffusion map. When processing ancient documents of this kind, the core technical challenge — known in academic circles as the "lacuna problem" — involves physically damaged characters, the high morphological variability of ancient languages, and cross-linguistic mixing (Hellenistic-period inscriptions, for instance, often blend Greek with local languages). Recent work such as DeepMind and the University of Oxford's Ithaca model (2022) has already demonstrated that AI can predict missing characters in ancient Greek inscriptions and infer their date and geographic origin, laying the technical groundwork for this type of cross-regional transmission analysis.

Reconstructing the Social Network of Oracle Visitors

The third application focuses on ancient Greek oracle sites, most famously the Oracle of Delphi. The Delphic Oracle was the most important religious and political center in ancient Greece, active from roughly the 8th century BCE to the 4th century CE — a span of over a thousand years. Envoys, kings, and commoners from Greek city-states, Persia, Egypt, and eventually Rome all made the journey, leaving behind dedication records, consultative inscriptions, and response texts that together constitute a massive dataset spanning the ancient Mediterranean world.

By reconstructing the social networks among these visitors, researchers can answer questions such as: Which city-states consulted the oracle most frequently? How did the concerns of different social classes differ? This is essentially a Social Network Analysis (SNA) of ancient society — except the data source is fragmentary ancient literature rather than modern social media.

Social Network Analysis (SNA) abstracts individuals (nodes) and relationships (edges) into graph structures. Originally developed in sociology and anthropology, it is now widely applied in internet research, financial risk control, and epidemiology. Applying SNA to ancient oracle records allows researchers to calculate the "centrality" of city-states, identifying which regions occupied hub positions in ancient information networks, or tracking unusual spikes in oracle consultations during specific political crises — in essence, using data science to profile ancient geopolitics.

Technical Significance: The "Reverse" Application of AI Reasoning

These three cases collectively point to a direction worth serious attention — deploying AI reasoning capabilities to "reconstruct the past" rather than "predict the future." This direction is coalescing into an emerging discipline: Computational History, which combines machine learning and natural language processing with traditional historical research methods to provide the humanities with an entirely new research infrastructure.

From a technical perspective, these tasks impose several distinctive requirements on AI:

1. Handling highly incomplete data: Ancient documents are frequently missing, damaged, or semantically ambiguous. AI must make reasonable inferences under conditions of incomplete information.
2. Cross-document correlation: Integrating fragments from different locations, eras, and languages into a coherent narrative.
3. Interpretability: Historical research cannot accept "black box" conclusions. Every association the AI surfaces must be traceable and verifiable by scholars.

By packaging these capabilities into reusable "Skills," Antigravity means that researchers in history, archaeology, and classical studies may soon be able to invoke powerful AI reasoning tools to support their work without needing deep programming expertise. Similar explorations — including Anthropic's Claude for Research and Microsoft's Copilot for Sciences — collectively point toward the democratization of AI tools.

The AI Opportunity for the Humanities

For a long time, AI's center of gravity has been commercial forecasting, code generation, and content creation. Antigravity's demonstration reminds us that the humanities offer equally fertile ground.

The core pain points of ancient history research — fragmented data, cross-linguistic and cross-regional scope, vast timescales — happen to be precisely the types of problems that modern AI handles well. When machine learning meets curse tablets, temple inscriptions, and oracle records, what we witness is not merely a technology demonstration, but a glimpse of a possible research paradigm: reconnecting historical clues that have been silent for two thousand years into meaningful narratives.

That said, these applications still warrant careful scrutiny. The associations AI surfaces are "hypotheses," not "conclusions," and the final historical judgment still requires professional review by human scholars. Research has shown that in practice with models like Ithaca, human-AI collaboration consistently outperforms either AI or human experts working independently — a principle that applies equally to Antigravity's historical reasoning scenarios. As a tool for accelerating research and expanding scholarly horizons, the potential of AI-driven historical reasoning is impossible to ignore. But its best role is to augment human scholarly judgment — not replace it.