Generative AI will attempt to reassemble, read, and restore the handwritten Latin texts of the Vindolanda tablets—fragile wooden documents from Roman Britain that are often too damaged or jumbled to decipher by eye. The problem is that many of these tablets remain unread, and manually recovering their texts is painstaking work. Earlier attempts to use computers for this task required disproportionate effort and were abandoned. Now, with vastly increased computing power and the success of generative AI on ancient Greek papyri, oracle bones, and cuneiform tablets, it is time to test whether the latest digital tools can succeed where earlier efforts stalled. If this research works, it could bring new texts to light from the Vindolanda collection—offering fresh insights into the Roman army’s daily life in Britannia. The improved AI models could then be applied to other handwritten objects across the British Museum collections, helping historians recover texts that would otherwise remain lost. This is fundamental, curiosity-driven research: it deepens our understanding of the past without an immediate practical application, but similar work on ancient materials has previously opened unexpected windows into history.
View original technical description
The Vindolanda tablets are an unrivalled documentary source for the Roman army and its activities in Britannia. The research will examine if Generative Artificial Intelligence (AI) can support the recovery of the handwritten Latin texts and in turn how this might be applied to aid the decipherment of other similar objects across the British Museum collections. Although much work has been carried out on the tablets, numerous texts are not deciphered, and the work to recover them is painstaking. A quarter of a century ago, research was undertaken to see if computers might help (Terras 2006 Image to Interpretation: An Intelligent System to Aid Historians in Reading the Vindolanda Texts), but the disproportionate effort required did not encourage development. Now massively increased computing power and the success of Generative AI in various ancient world applications (Sommerschield et al. 2023 ‘Machine Learning for ancient languages: A survey’, Computational Linguistics, 49.3) mean that it is time to test the latest digital tools (e.g. GPT-4o) to undertake four main tasks: identify and recognise handwritten characters; digitally reassemble fragments; restore missing text when objects are fragmentary or text has been lost; decipher palimpsest (overlying) texts. These tasks would be undertaken using Generative AI pipelines that have already been created for Greek papyri (e.g. Swindall), oracle bones (e.g. Zhang), and cuneiform tablets (e.g. Dahl) and have the potential to bring new texts to light from the extraordinary Vindolanda collection, and to create improved models which could be applied to other BM collections.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know