The Machine That Reads Missing Words

How Ithaca helps historians test the possibilities hidden in damaged inscriptions

History often survives in incomplete sentences. Weather, fracture and reuse can erase the words that located an inscription in a particular city or century. Reconstructing the text therefore requires more than recognising letters; historians must compare language and material context across many surviving examples.

Ithaca offers a new way to organise that comparison. Developed by historian Thea Sommerschield and AI researcher Yannis Assael, the neural network learned patterns from roughly 80,000 Greek inscriptions. Given a damaged passage, it proposes restorations and estimates where and when the text was produced.

Its predictions are not final translations. Each suggestion is a hypothesis that specialists must judge against evidence the model may not capture, including an artefact’s physical setting or unusual local usage. The system can scan relationships across a corpus far faster than one researcher.

The strongest results came when human expertise and machine prediction were combined. Ithaca suggests a practical role for AI in historical research: a tool that expands the possibilities scholars can examine rather than an automated authority on the past. The missing words still demand interpretation; the machine helps historians decide where to look.

Video by Nature

Producer: Shamini Bundell