Paul Dilley, the Erling B. “Jack” Holtsmark associate professor in the Classics and department executive officer in the Department of Religious Studies, has received a $500,000 grant from Schmidt Sciences through the Humanities and AI Virtual Institute (HAVI) to develop artificial intelligence models designed to restore and interpret damaged ancient manuscripts.
The three-year project, titled “AI Models with Reinforcement Learning to Edit Text in Damaged Manuscripts,” will produce software capable of reconstructing missing or illegible sections of historical documents, allowing scholars to study texts that have remained inaccessible for decades—or even centuries.
Addressing a fundamental challenge in the study of ancient texts
The vast majority of ancient writings have not survived, perishing sometimes due to benign neglect, but also because of targeted destruction, and occasionally even through natural disasters, from floods to volcanic eruptions. A few texts facing these events have been miraculously preserved, kept safe under sands or in carbonized form, but remain significantly deteriorated. Although advanced imaging technologies have helped reveal previously hidden writing, they cannot fully reconstruct texts when letters, words, or entire passages have been lost.
Dilley and his collaborators aim to address limitations by combining artificial intelligence with traditional methods of scholarly textual editing.
The international research team* will develop AI models trained with reinforcement learning to analyze patterns of manuscript damage, reconstruct faded or fragmentary letters, and propose restorations for missing words and sentences.
By integrating human expertise in ancient languages and historical context with computational techniques, the system will generate and evaluate multiple possible reconstructions, helping scholars restore texts more efficiently and systematically.
An initiative with broad potential
Dilley emphasized the broader potential of the initiative, particularly in unlocking texts long thought lost: “This is an amazing opportunity to leverage both enhanced imaging techniques and cutting-edge AI training. Our project aims to take a major step forward toward recovering some of the most exciting but currently obscured texts, like the scriptures of the Manichaeans, an early Christian arch-heresy, from Coptic codices, and the philosophical writings of the Epicureans, preserved in the Herculaneum scrolls.”
This funding from Schmidt Sciences will support the development of specialized software tools and an interactive user interface that allows scholars to train AI systems through direct engagement with manuscript images and texts. The project will also establish benchmarks for evaluating AI-assisted restoration methods, bringing together best practices from computer science and the long-standing humanistic tradition of textual scholarship.
Beyond advancing research, the project will create new opportunities for students and scholars to engage with emerging AI methods in the humanities. As the models are developed and refined, the team hopes to expand their application to additional damaged manuscript collections, opening new possibilities for recovering and studying lost portions of human history.
*Paul Dilley's co-principal investigators include James H. Brusuelas of the University of Kentucky, Vincent Christlein of Friedrich-Alexander University Erlangen–Nürnberg, Gianluca Del Mastro, an independent scholar, and Federica Nicolardi of the University of Naples Federico II.