New AI Model 'Apollo' Reconstructs Missing Passages in 2,000-Year-Old Greek Texts

Classicists facing a manuscript missing half its words usually confront a puzzle with no definitive solution, requiring weeks of painstaking cross-referencing just to guess at a single torn sentence. A newly released artificial intelligence model, Apollo, is breaking that bottleneck by filling in the blanks of 2,000-year-old texts in seconds.[1][2]

Unveiled on September 23, 2026, in Vienna, Apollo is the world's first large language model built specifically for Ancient Greek. The system, funded with approximately €400,000, is the result of a collaboration between the Austrian Academy of Sciences (ÖAW), French AI developer Mistral AI, and technology firm Sail Reply. Unlike general-purpose chatbots, Apollo was trained on a specialized corpus of roughly 600 million words drawn from historical manuscripts, papyri, and stone inscriptions.[1][4][5]

Ancient Greek writing lacks the word spaces and punctuation that modern readers rely on, making damaged texts exceptionally difficult to parse. Building on Mistral Small—a foundation model with 24 billion parameters—Apollo examines the surviving characters on either side of a gap and uses a 'fill-in-the-middle' approach to predict what originally appeared there without requiring an exact character count.[3][5]

In blind tests involving deliberately concealed passages on original papyri, the model's accuracy outpaced earlier restoration systems. For gaps of up to 10 characters, Apollo included the correct reading among its top 20 suggestions in 80.6 percent of documentary papyrus cases, 54.6 percent of literary texts, and 61 percent of stone inscriptions.[3][4]

The scale of the problem the model addresses is vast. According to the Austrian Academy of Sciences, 92 percent of the world's surviving papyri—more than 1 million fragments held in museums and libraries—remain undeciphered. 'Apollo is taking the study of historical manuscripts and documents into a new era,' said Heinz Faßmann, President of the ÖAW. 'The use of this new language model will open up entirely new research questions and expand our historical knowledge.'[1][4]

The system has already been tested on some of antiquity's most notoriously difficult documents. Researchers applied Apollo to P.Herc. 1667, a carbonized philosophical scroll recovered from the volcanic ruins of Herculaneum after the Mount Vesuvius eruption in 79 A.D. The model successfully proposed an alternative reading for a damaged passage previously interpreted as 'to do as is in our nature,' demonstrating its ability to handle complex Stoic philosophy.[3]

Because the Greek language evolved significantly over centuries, Apollo was designed to recognize and adapt to regional dialects and historical periods. Project lead Anna Dolganov noted that the system can seamlessly switch to Homeric Greek when analyzing epic poetry, or employ the Doric dialect when reconstructing a regional stone inscription. This contextual awareness helps scholars narrow down the date and origin of undocumented fragments.[5]

Crucially, the developers designed Apollo to assist human experts rather than replace them. Instead of generating a single, authoritative historical verdict, the model provides a shortlist of statistically probable words or phrases for a scholar to evaluate. This design choice prevents probability-based AI output from silently rewriting the historical record, ensuring that human judgment remains central to the reconstruction process.[2][5]

The first version of Apollo is now freely available to academic researchers through a chatbot interface. As scholars begin feeding it administrative records, tax receipts, and private letters from the Roman and Hellenistic periods, the next phase of discovery will not come from unearthing new ruins, but from finally reading the archives we already have.[2][4]