Hundreds of thousands of Ancient Greek papyrus fragments sit in academic collections, many with words and passages missing. A new language model called Apollo is intended to help scholars reconstruct the material faster without handing final interpretation to a machine.
The Austrian Academy of Sciences is releasing Apollo in partnership with French AI laboratory Mistral and technology services company Sail Reply. The organizations describe it as the first advanced large language model for Ancient Greek. It was trained on roughly 600 million historical Greek words taken from manuscripts, papyri and inscriptions.
A Specialist Model for Damaged Texts
Apollo will be freely available to academics through a chatbot interface. Researchers can use it to locate fragments relevant to their specialties, identify promising lines of inquiry and generate statistically likely candidates for text missing from damaged records.
Reconstructing a papyrus normally demands several kinds of expertise. Ancient Greek was written without spaces between words, so a scholar must first identify word boundaries. The document must then be dated and interpreted within its social and political setting before reference works can help narrow possible missing language.
Apollo compresses much of that specialist context into one model. Austrian Academy historian and papyrologist Anna Dolganov said the system changes its suggestions when it encounters Homeric Greek or an inscription written in the Doric dialect. The aim is not merely to predict common words, but to respond to the period, form and linguistic variety visible in a fragment.
For scholars doing repetitive reconstruction work, several plausible options could accelerate the first stage of analysis. Oxford classics professor Armand D’Angour said receiving three likely words for a gap would speed the process considerably.
The Likely Discoveries Are Small but Numerous
Apollo is not expected to uncover a lost collection of famous plays or overturn the broad outline of the ancient world. University College London professor Stephen Colvin noted that many unrestored papyri are mundane records such as personal letters, marriage contracts and civil-service documents.
That ordinary material can still add important detail. A completed letter or administrative record may support an existing theory or reveal a small fact about daily life. D’Angour described each restoration as adding another element of knowledge rather than producing a spectacular single discovery.
The system may also help academics search large collections more efficiently. A specialist could use the model to surface fragments connected to a narrow topic instead of examining records one by one. If the approach works, Sail Reply partner Dimitris Vlitas believes similar methods could be applied to Latin, Egyptian or other fields built around large historical corpora.
Those possibilities are projections from the model’s developers and outside scholars. WIRED does not report a completed independent benchmark showing how often Apollo’s restorations match expert conclusions.
Scholars Retain the Final Historical Judgment
Language models generate probabilities, creating an obvious danger when they are asked to repair the historical record. A confident but incorrect completion could turn a modern statistical guess into an apparent ancient statement.
Apollo is designed to reduce that risk by offering several candidate words rather than silently producing a single definitive reconstruction. A qualified scholar must choose between the suggestions and evaluate them against the document’s physical and historical evidence.
Dolganov said human competence must remain central and warned against complete dependence on AI transcription or interpretation. That boundary defines Apollo’s intended role: it can index knowledge, propose options and reduce painstaking search, but it does not authenticate its own answers.
The model’s release will show whether a narrowly trained AI can become a practical research instrument for classicists. Its success will depend less on fluent Ancient Greek output than on whether experts can use its suggestions without allowing probability to harden into false history.
