Researchers have developed the world’s first large language model dedicated to Ancient Greek, using artificial intelligence to reconstruct missing passages in historical documents and potentially uncover new evidence about life in the ancient world.
The AI system, named Apollo, was developed by the Austrian Academy of Sciences (ÖAW) in collaboration with European AI companies Mistral AI and SAIL Reply. Unveiled on 23 September 2026, the technology is designed to help archaeologists and historians decipher fragmentary texts that have remained incomplete for centuries.
Trained on approximately 600 million words of Ancient Greek, Apollo can analyse surviving passages and suggest plausible reconstructions of missing words, sentences and longer sections in seconds.
The first version of the system is now freely available to researchers and the public, opening up new possibilities for studying the vast collections of ancient documents held in museums and libraries worldwide.
AI tackles one of archaeology’s biggest challenges
Hundreds of thousands of ancient inscriptions and more than a million papyrus fragments survive in collections around the world. However, much of this material is damaged, incomplete or difficult to decipher.
According to the Austrian Academy of Sciences, 92 per cent of papyri still await decipherment, representing an enormous body of historical evidence that has yet to be fully explored.
Reconstructing these documents traditionally requires specialist knowledge of ancient languages, historical context and the writing conventions of different periods. Even experienced scholars can spend considerable time attempting to restore a few missing words.
Apollo aims to accelerate this process by analysing the surviving text and generating possible reconstructions that researchers can assess against the original material.
Unlike a general-purpose AI chatbot, the system has been developed specifically for Ancient Greek, enabling it to recognise linguistic patterns and suggest additions that reflect the vocabulary, grammar and style of historical texts.
Apollo achieves 80 per cent accuracy in reconstruction tests
Despite being trained on a relatively small dataset by modern AI standards, Apollo has demonstrated promising results in reconstructing damaged texts.
While conventional large language models are often trained on tens of billions of words, Apollo was developed using approximately 600 million words of Ancient Greek.
In tests involving original papyrus texts with deliberately concealed passages, the model correctly reconstructed around 80 per cent of the missing content.
“With an existing original text on papyrus, the model achieves a hit rate of around 80 percent for masked, i.e., deliberately hidden, passages,” explained Anna Dolganov, a researcher at the Austrian Archaeological Institute of the ÖAW who led the project.
The system was also assessed in a blind study involving international experts, who compared its suggestions with reconstructions produced by human scholars.
The experts rated Apollo’s proposed additions as at least as good as human-generated reconstructions in 77 per cent of cases. In 16 to 20 per cent of cases, they considered the AI’s suggestions better.
These results suggest that specialist language models could become valuable research tools, particularly when scholars are confronted with damaged texts containing several possible interpretations.
From Homer to Roman Egypt: AI recognises different forms of Greek
One of Apollo’s most significant capabilities is its ability to distinguish between different periods and styles of Ancient Greek.
The language evolved considerably over the centuries, and the vocabulary and grammatical conventions of Homeric poetry differ substantially from those found in documents produced under Roman rule.
According to Dolganov, Apollo can identify these distinctions using contextual clues such as sentence structure and vocabulary.
“Our LLM can draw far-reaching conclusions with comparatively little knowledge about a context,” she explained.
For example, the model can recognise the linguistic register associated with Homer’s Odyssey and generate suggested additions that reflect its distinctive style.
This ability is particularly important when reconstructing documents whose date, origin or literary context may be uncertain.
By producing suggestions appropriate to the historical period, Apollo could help researchers narrow down possible interpretations of incomplete texts.
Ancient birth record and Vesuvius scroll reconstructed
During the system’s initial presentation, Dolganov demonstrated Apollo’s capabilities using three historical examples.
One involved a Greek birth announcement from Roman Egypt dating to AD 149. The AI successfully reconstructed the missing portions of the document, illustrating its potential for restoring administrative records from antiquity.
Greek birth announcement from Roman Egypt, dated to AD 149. Image: Austrian National Library.
Apollo was also used to reconstruct important sections of a scroll charred during the eruption of Mount Vesuvius in AD 79.
The eruption, which destroyed Pompeii and Herculaneum, preserved a collection of ancient scrolls in a heavily carbonised state. Their fragile condition has made them exceptionally difficult to study.
The AI demonstration showed how reconstructed passages could help researchers interpret text recovered from such damaged material.
A third example involved a fragmentary inscription from the Black Sea region. Its reconstruction provided evidence that Roman law was already operating in a particular area, a finding further supported by evidence of a Roman tax on prostitution.
The examples illustrate how restoring even a small section of a damaged document can have wider implications for understanding ancient administration, legal systems and economic activity.
New discoveries in ancient social and economic history
Researchers believe Apollo could be particularly valuable for studying everyday life in antiquity, an area in which historical evidence is often scattered across thousands of surviving documents.
While famous literary works have received extensive scholarly attention, many administrative records, tax receipts, contracts and private letters remain unpublished or incompletely understood.
Individually, these documents may appear unremarkable. Studied collectively, however, they can reveal important patterns in trade, taxation, employment and social relationships.
“With the help of AI, a new picture of ancient social and economic history is emerging, as many insights only become apparent from the sheer volume of documents – for example, tax receipts,” Dolganov said.
By accelerating the reconstruction of damaged texts, Apollo could make it possible for scholars to examine much larger collections of historical material.
It could also generate alternative interpretations that researchers might otherwise overlook.
Nevertheless, Dolganov emphasised that the system is intended to support scholarly investigation rather than replace human expertise.
“Apollo doesn't just accelerate work, it provides genuine inspiration,” she said. “The model suggests additions to texts that one might otherwise never have thought of, or only after a long search.”
Researchers must still evaluate the AI’s suggestions against the surviving evidence and establish whether a proposed reconstruction is historically and linguistically plausible.
Researchers plan to develop an AI for Latin
The newly released version of Apollo represents the first stage of a wider research programme.
Future developments are expected to combine the reconstruction of missing passages with the ability to decipher previously unexplored papyri.
The research team also plans to introduce thematic and semantic search capabilities, allowing scholars to search across the Ancient Greek textual tradition for related documents, parallel passages and supporting evidence.
Such tools could help researchers identify connections between texts that were produced centuries apart or survive in different collections.
The Austrian Academy of Sciences and its partners also intend to extend the technology to Latin.
A Latin-language model would have access to a substantially larger body of training material, with the project anticipating a corpus of approximately 12 billion words.
The researchers have yet to choose a name for Apollo’s Latin counterpart.
European collaboration brings ancient history and AI together
Apollo was developed with funding of approximately €400,000 through the collaboration between the Austrian Academy of Sciences, Mistral AI and SAIL Reply.
The system is owned by the ÖAW and is available as a free, open-access application.
Heinz Faßmann, President of the Austrian Academy of Sciences, described the project as an important development for historical research.
“Apollo ushers in a new era for the study of historical writings and documents,” he said.
Guillaume Lample, co-founder and chief scientist of Mistral AI, highlighted the potential of developing AI systems specifically for scientific research.
“By helping to reconstruct texts that have only survived fragmentarily for two thousand years, Mistral’s technology opens a new window to our shared cultural heritage,” he said.
The project takes its name from Apollo, the ancient Greek god associated with the arts, knowledge and prophecy.
With its first version now available, the technology offers researchers a new way to investigate the surviving written record of antiquity. Its wider significance will depend on how effectively scholars can use its reconstructions to verify historical evidence and explore documents that have remained unread for centuries.
The Apollo AI system is available at apollo.vbc.ac.at
Sources : OAW