AI as an Ancillary Science of History? Reading Texts and Mapping Scribal Hands in the Dunhuang Manuscripts

Discovered in 1900 in northwestern China, the Dunhuang manuscripts constitute an exceptionally rich record of the written cultures of medieval China and Central Asia. Comprising nearly 70,000 written artefacts dating from the fifth to the eleventh centuries, the corpus includes Buddhist scriptures, administrative records, literary and didactic texts, and medical and divinatory materials. Its diversity offers considerable opportunities for historical inquiry, but its scale and dispersion across collections complicate systematic study. Drawing on work undertaken in collaboration with the Bibliothèque nationale de France and the British Library, this presentation explores two complementary approaches to reading and relating these manuscripts.

We first address the challenges of Chinese handwritten text recognition, particularly the uneven frequency of characters and the difficulty of obtaining sufficient training examples for rare forms. By automatically aligning existing digital transcriptions with images of historical documents, we assembled a training dataset of ten million lines without new manual annotation. Models trained through this approach achieve character error rates below 1% on several evaluation datasets.

We then turn from textual content to scribal hands. Using self-supervised representation learning, we learn visual representations of handwriting without supplying writer identities during training. Evaluated on a corpus of signed ninth-century manuscripts, these representations support the retrieval of manuscripts with similar handwriting and the mapping of affinities between hands. We examine how these affinities can be interpreted in palaeographical terms and what further evidence is required to establish a scribal attribution.

Together, these approaches invite reflection on AI’s place among the ancillary sciences of history. By building on existing scholarship and learning from manuscript images, they extend the possibilities for reading and comparing sources at scale.