A broken line on an ancient papyrus can leave historians facing a maddening puzzle. The surviving words may describe a tax, a family dispute or an instruction to a public official, but the missing portion determines what the writer actually meant. Researchers in Austria have developed a specialised model to suggest possible restorations, an example of how AI reads Ancient Greek without pretending to know exactly what time has erased.
The system, called Apollo Restore, comes from the Austrian Academy of Sciences' Decoding Antiquity initiative. In a September 2026 announcement, the academy described a model trained on a large body of historical Greek writing and designed to offer ranked alternatives for lost passages. Those alternatives are proposals for scholars to investigate, not authenticated transcripts of words that have vanished. That distinction is central to the project's promise.
A small gap can change a very old story
Ancient history often survives in fragments rather than neatly bound books. A letter might have lost its opening, a legal document its crucial condition or an account its final amount. Sometimes only the edges of a sentence remain. The historian must consider the words that are visible, the language used in that period and the kinds of documents likely to have been written in that place.
This is slow, specialised work. Ancient Greek changed across centuries and communities, and surviving material ranges from formal literary prose to the hurried handwriting of ordinary administration. Even identifying where a missing word would begin can be difficult. A plausible restoration has to fit the grammar and the context. It should also make historical sense without being chosen merely because it produces the most entertaining interpretation.
A good reconstruction can open a window onto everyday life. Taxes, wages, marriages, property, trade and personal quarrels leave traces in letters and records. What sounds like a dry exercise in filling gaps may alter how historians understand the lives of people whose experiences were rarely recorded in grand monuments. It is precisely because these documents are so valuable that inserting the wrong words carries a risk.
How AI reads Ancient Greek without inventing history
The researchers have developed Apollo Restore as a language model for historical Greek, rather than relying on an ordinary consumer chatbot. The project says its training material amounts to approximately 600 million words drawn from premodern and early modern Greek sources. The model can propose possible missing text, including gaps whose length is uncertain, and rank the candidates according to its learned patterns.
The attraction is obvious. A specialist looking at an incomplete sentence can compare more possibilities than might readily come to mind. The system can act like an unusually well-read assistant that never tires of searching for related expressions. It may reveal a grammatical construction or parallel phrase that helps a researcher reconsider an initial interpretation.
Yet language probability is not a historical witness. A convincing line might be something a Greek writer commonly could have said, rather than what this particular person actually wrote. An author might have used an unusual name, made an error or described an event for which few parallels survive. A system trained to find the most plausible continuation can easily smooth away precisely the distinctive evidence a historian wants to preserve.
That is why the strongest use is to present alternatives with clear uncertainty, not silently replace damage with fluent text. The project's research description places particular emphasis on separating recorded readings from reconstructions and making uncertainty visible. If readers cannot tell original letters from machine-suggested ones, the technology risks changing historical evidence while appearing to restore it.
How encouraging are the test results?
The academy reports strong experimental performance, including an accuracy figure of about 80% on a held-out evaluation involving masked passages. Its announcement also describes a blinded expert comparison in which the system's proposals were judged at least as good as human restorations in 77% of the evaluated cases. Those are meaningful results for the researchers' defined tests, but they require interpretation.
Masking words in a text whose original version is known creates a useful benchmark: the correct answer exists somewhere to be checked. A truly damaged manuscript is harder. Scholars may not know exactly how much is missing, the text may contain peculiar spelling and crucial context may have disappeared. Success at recovering deliberately removed passages does not prove the same reliability with every ancient fragment.
The comparison with experts is similarly helpful without being conclusive. A panel may judge one proposal more plausible than another yet have no independent way of knowing what an unknown scribe wrote. The result measures the quality of possible reconstructions within an evaluation procedure. It is not evidence that a computer has recovered the original historical truth in three-quarters of real cases.
For readers, the most important number may therefore be the one that cannot yet be supplied: how many previously undeciphered or uncertain documents will ultimately be clarified to historians' satisfaction after a transparent, independently reviewed process? That is the long-term test, rather than a single score achieved on selected examples.
The difference between reading ink and imagining missing ink
Apollo Restore is principally addressing textual reconstruction. Another major problem is turning faint, damaged handwriting into reliable letters in the first place. A photograph of a papyrus may contain tears, stains, shadows and almost invisible strokes. An image-recognition model must first decide what marks are really present before a language model can offer a reasonable interpretation of an incomplete line.
The research programme is developing visual systems for that task as well. It hopes to work with page images and historical language knowledge together, rather than making specialists feed a perfect transcript into a model. Combining these capabilities could make a large archive easier to search and compare. But it also introduces another possible source of error: if the visual system mistakes a letter, the reconstruction model may build an elegant but false sentence around it.
This is not quite the same problem described in LiveAIWire's report on AI-assisted work with sealed Vesuvius scrolls. In that case, researchers have been trying to distinguish physical writing hidden inside damaged material. Apollo's focus is helping to reason about missing or fragmentary text once historical language is available. Both efforts show how AI can extend the reach of human scholarship, while answering very different questions.
The overlooked archives may hold the richest rewards
The academy says much of the surviving documentary record remains unread, unpublished or only partly accessible. That is a labour problem as much as a technology problem. Experts must travel, catalogue, decipher, compare and edit material before others can use it. A system that provides sensible starting points could make those hours more productive, even when its suggestions ultimately prove wrong.
There is precedent for digital tools changing what scholars can attempt. AI-assisted archaeology has helped researchers search large spaces for clues that would be difficult to examine individually. Work on the Nazca landscape demonstrates a different use, where machine analysis helps direct attention to physical traces. In textual history, the equivalent opportunity is finding documents and patterns that would otherwise remain buried in collections rather than sand.
However, scaling up can create a subtle temptation to publish machine-generated interpretations faster than they can be checked. A searchable database filled with fluent guesses might look impressively complete and still mislead future researchers. Good digital editions therefore need provenance: images of the original material where possible, notes explaining uncertainty, and a clear record of human editorial decisions.
A useful archive is not the one with the fewest blank spaces. It is the one in which readers can see which spaces are genuinely known and which remain open to interpretation. A careful Apollo workflow might preserve multiple reconstructions rather than force one preferred answer into the historical record.
History is allowed to remain uncertain
It is easy to picture the technology as a machine giving lost people their voices back. The more responsible image is a scholar receiving a set of better questions. Does this phrase match the writer's dialect? Is there a comparable tax document? Could the apparently missing name be something less familiar? Why has the system ranked one version above another?
That approach respects both the model's strengths and its limits. Pattern recognition can direct attention and reduce routine work; historians supply the expertise to judge whether a suggestion is warranted. As evidence improves, their conclusions may change. There is no need to treat uncertainty as failure when uncertainty is the honest state of the record.
Apollo could help researchers reach parts of the ancient world that have remained frustratingly inaccessible. Its real contribution will not be a perfectly completed manuscript with every hole concealed. It will be a better, more transparent way to investigate the words that have survived and to ask what might once have stood between them.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.