Apollo brings probabilistic AI to damaged Greek papyri

What Apollo actually does

Ancient Greek manuscripts often have no spaces between words, and damaged papyri can leave scholars unsure where one word ends and another begins. Reconstructing a passage typically requires several kinds of expertise:

Apollo is designed to bring those steps into one system. It can search for papyrus fragments related to a research topic, suggest promising areas for study, and propose statistically likely words or passages for gaps.

The model is intended to account for differences within the language. Anna Dolganov, a historian and papyrologist at the Austrian Academy of Sciences, says it can complete Homeric text in Homeric Greek and use the Doric dialect when working with a Doric inscription.

Dimitris Vlitas, a Sail Reply partner, told WIRED that a year ago, extracting knowledge this way seemed unimaginable.

The value is in the shortlist

The system’s most practical output may be a small set of candidates rather than a confident-looking answer. Armand D’Angour, a professor of classics and literature at the University of Oxford, says that if a machine offered three words suitable for a particular gap, it would make the reconstruction process much easier.

That matters because the field has very few people with deep enough knowledge of Ancient Greek history to handle every damaged text. Apollo could reduce the time spent determining what a document says and leave researchers more time to study what it means.

My read is that this makes Apollo a research assistant, not an automated archaeologist. Its promise lies in compressing the tedious part of expert work while keeping the expert responsible for the choice.

The announcement is quiet about the metric that would make this promise testable: how often Apollo’s suggestions match later scholarly judgments. Without that figure, the 600-million-word training corpus describes the model’s scale, not its accuracy.

More everyday history, not lost masterpieces

Apollo is unlikely to rewrite the broad picture of the ancient world. Many unreconstructed papyri are ordinary documents—personal letters, marriage contracts and records produced by government officials.

Steven Colvin, a professor of classics and historical linguistics at University College London, cautions against expecting the sudden discovery of several new plays by Sophocles. The more plausible payoff is a larger number of small findings: details about daily life and evidence that supports existing historical hypotheses.

That incremental value is still significant. D’Angour argues that every reconstructed text adds another small piece to the record of the ancient world.

The risk is misplaced confidence

A language model works with probabilities, which makes it dangerous when its output enters historical records. Apollo is meant to reduce that risk by presenting several possible words instead of a single replacement. Dolganov says the scholar’s expertise must remain central.

If the approach works, Vlitas says it could extend to other ancient languages, including Latin and Egyptian, and to disciplines where large collections of material need to be processed and indexed. The same pattern has already appeared elsewhere in science: OpenAI said its models solved a 200-year-old mathematics problem, while Google DeepMind released a large dataset on how genetic mutations affect molecular biology that it assembled with AI.

The tension is straightforward: Apollo can make obscure documents easier to approach, but the more fluent its reconstructions become, the more important it is that researchers treat them as proposals rather than recovered facts.