A gap in an ancient inscription is not an empty field in a modern document. A missing name may appear in a dialect known from only a small number of other texts. Letter shapes may reflect a local alphabet. The stone may preserve enough grammar to constrain a restoration, but not enough to identify one unique answer. In this setting, the apparently simple task of asking AI to fill in the blank becomes a demanding test of what language models actually know.[1][2]
Researchers linked to the universities of Udine and Ca’ Foscari Venice describe this challenge in the LACUNAE project, published in Digital Scholarship in the Humanities on 12 October. Their work investigates natural-language processing tools for restoring fragments of archaic and classical Greek epigraphic texts, with special attention to the dialectal and orthographic variety that broad models can flatten away.[1][2]
The project examines established approaches including Ithaca and Ancient Greek BERT, and considers whether general-purpose and instruction-tuned large language models can contribute knowledge not readily available in the smaller epigraphic corpus. Its stated aim is not to declare the best model from a single spectacular restoration. It is to evaluate approaches across dialects, different degrees of missing text and examples unseen during training.[1][2]
Ancient writing is an evidence problem, not autocomplete
Imagine an inscription with part of a civic office holder’s title missing. A general text model might favour a common Attic Greek construction because it is more familiar from its training sources. But an unusual regional formula may be the historically plausible one. The most statistically probable continuation is not automatically the best reconstruction.[1][2]
A historian also has to distinguish the surface that survives from the proposed addition. Even a grammatically attractive completion can be wrong if the number of missing characters does not fit, the letter forms belong to another period, or an institution named in the suggestion did not exist in the relevant place. Those restrictions are not optional decorations; they are part of the claim.[1][2]
The evidence a restoration must reconcile
| Evidence layer | Why it matters |
|---|---|
| Physical inscription | Visible marks, line spacing and damage restrict possible lettersSource 1 |
| Local dialect and alphabet | A statistically common spelling may be locally or historically wrongSource 1 |
| Parallel epigraphic texts | Similar dedications or civic formulae offer comparators, not automatic proofSource 1 |
| Literary sources | Broader phrasing may inform a hypothesis but differ in registerSource 1 |
| Model output | A proposal to assess and document, not newly recovered primary evidenceSource 1 |
The term 'lacuna' describes an absence in the evidence. Its meaning is a useful warning against treating generated prose as retrieved fact. A responsible digital edition should preserve the uncertain status of conjectural letters and, ideally, provide alternative proposals and reasons for choosing one. The model is an assistant to that process, not a witness to the destroyed inscription.[1][2]
Why dialect and alphabet create a difficult test
Most modern language models inherit a strong statistical advantage from abundant languages and standardized spelling. Archaic Greek inscriptions violate both conditions. Their surviving record is uneven, signs can differ between cities and time periods, and local forms may not correspond neatly to later classical spellings.[1][2]
Data scarcity creates a second difficulty. Splitting a tiny corpus into training and evaluation sets can leave models with very little evidence of a rare construction. A high score on material resembling training examples might conceal a failure precisely where historians most need assistance: unfamiliar fragments. The LACUNAE paper discusses testing on unseen material and across dialects, though its accessible abstract does not provide a general numeric accuracy result that can responsibly be quoted for the whole project.[1][2]
The researchers also want literary sources to inform restoration where appropriate. That is valuable because a phrase might survive in a poem or historical account even if few inscriptions preserve it. But literary language differs from the formulae used in public decrees, dedications and funerary stones. Broader context can guide a hypothesis without settling it.[1][2]
What a useful research tool would show
A genuinely valuable interface would reveal which part of a suggestion came from the original stone, which from comparable inscriptions, which from literary parallels and which from a model’s statistical inference. A specialist could then accept, revise or reject the proposal with an auditable trail. A polished sentence alone would make the scholarly task harder by disguising the boundary between evidence and invention.[1][2]
The LACUNAE team has linked code and datasets from the project. That allows other researchers to inspect and test methods, but it is not a claim that every proposed reconstruction has passed expert review or that the scarce-data problem has been solved. Full-paper access restrictions also limit independent assessment of detailed results from the public abstract alone.[1][2]
The deeper promise is not the sensational idea that AI has 'decoded' a lost civilization. It is that computational methods might help scholars search more systematically through constrained alternatives while keeping uncertainty visible. Ancient history can benefit from faster hypotheses. It cannot dispense with the discipline of proving why one hypothesis is better than another.[1][2]
Sources & further reading
- The LACUNAE project: evaluating NLP models for the restoration of archaic Greek inscriptions Zunino and colleagues, Digital Scholarship in the Humanities · 12 Oct 2026 · Accessed 2 Nov 2026
- The LACUNAE Project: Performing AI on Archaic Greek Epigraphy University of Udine research repository · Accessed 2 Nov 2026