Researchers have built the first advanced language model for Ancient Greek to help scholars reconstruct hundreds of thousands of damaged papyrus fragments sitting in academic libraries worldwide. The Austrian Academy of Science released Apollo on Wednesday, developed alongside French AI lab Mistral and technology services company Sail Reply. The tool aims to speed up the painstaking work of filling in missing words and phrases in tattered historical documents, a task that currently requires rare expertise and meticulous manual effort.
Apollo was trained on roughly 600 million historical Greek words pulled from manuscripts, papyri, and inscriptions. The model will be available for free to academics through a chatbot interface. It's designed to help scholars quickly identify papyrus fragments relevant to their specialized fields and suggest promising new research directions. When documents are damaged, the system fills in gaps with the most statistically probable words or passages based on the context it recognizes.
The model adapts its language to match what it encounters, according to Anna Dolganov, a historian and papyrologist at the Austrian Academy of Science. When Apollo sees Homer, it supplements with Homeric Greek; when it encounters an inscription in Doric dialect, it uses that dialect. Dimitris Vlitas, partner at Sail Reply, says that unlocking knowledge this way "was unthinkable a year ago." The report notes that restoring papyrus traditionally requires an academic to identify word divisions in Ancient Greek writing (which has no gaps), accurately date the document, consider socio-political contexts, and consult reference materials to choose appropriate words for missing sections.
The tool probably won't reshape broad understanding of the ancient world, since many unrestored papyri are mundane documents like personal letters, marriage contracts, and civil service papers. But it could uncover new details about antiquity and confirm existing scholarly theories. Armand D'Angour, a professor at Oxford (home to the world's largest ancient papyrus collection), calls it "very exciting" and says having a machine suggest three possible words for a gap would speed up work considerably. To prevent errors from polluting the historical record, Apollo proposes multiple word options for scholars to choose between rather than auto-filling text. If successful, the same approach could be applied to other ancient languages like Latin or Egyptian, or any academic field that would benefit from distilling and indexing large amounts of material. The crucial point, Dolganov emphasizes, is that human competence must remain—problems arise when scholars become totally reliant on AI transcriptions and interpretations. The model represents a tool for acceleration, not replacement, allowing researchers to spend less time deciphering what documents say and more time analyzing what they mean. Language models have already shown success in specialized academic tasks, from solving centuries-old math problems to mapping how genetic mutations affect molecular biology, suggesting AI's role in scholarship will continue to expand. Apollo's release marks another step toward AI serving as a research accelerant in fields where expertise is scarce and source material is vast.