A new AI model is trying to guess what's missing from ruined ancient Greek manuscripts.
Researchers built a large language model trained to fill in the gaps in papyrus fragments, the brittle, incomplete records that make up much of what survives from antiquity. Many ancient Greek texts exist only as scraps, torn and faded, with holes and missing lines breaking up the original writing. The model is designed to predict the words most likely missing from those gaps, based on patterns learned from the surrounding text. The aim, according to the researchers, is to recover new details about ancient life that have been lost to the damage.
Papyrus fragments are among the few direct windows into everyday antiquity, but scholars have spent careers staring at gaps they can't fill by hand. If a model can make credible guesses at scale, it could speed up reconstruction work that used to take specialists years, fragment by fragment.
The catch is the one that dogs every language model: a fluent guess about missing text is not the same as recovering what was actually written. Historians will need to treat its output as a hypothesis worth checking, not a transcription to trust.