Can AI Read Coptic Ostraca?: Formula, Fragment, and the Reconstruction of Meaning in O.Frangé 1–27

Presentation slides for “Can AI Read Coptic Ostraca? Formula, Fragment, and the Reconstruction of Meaning in O.Frangé 1–27,” presented at the online mini-conference AI for Ancient Studies (AI4AS), held as part of DH2026 (27–31 July 2026, Daejeon, South Korea). This study examines how three large language models, GPT-5.5 (accessed through ChatGPT), Gemini 3.5 Flash, and Claude Sonnet 5, translate Coptic documentary texts. The experiments were conducted through the services’ free-tier interfaces in July 2026, without task-specific fine-tuning or supplementary contextual information supplied by the researcher.

Drawing on O.Frangé 1–27, a set of Coptic monastic letters from late antique Western Thebes, the presentation focuses on nine selected documents. AI-generated translations are compared with the published French translations, with particular attention to practical requests, epistolary formulas, fragmentary passages, and scriptural or religious language. Recurring translation errors and interpretive choices are analyzed as evidence of how LLMs reconstruct meaning.

The analysis highlights four tendencies: completion bias, which turns incomplete passages into coherent statements; semantic inflation, which adds emotional or interpretive weight to conventional expressions; genre distortion, which recasts documentary communications as personal or devotional reflections; and the erasure of formulaic structure, which obscures repetition and epistolary conventions. The presentation argues that evaluating LLM translation in papyrology requires attention not only to linguistic accuracy but also to interpretive behavior, especially the handling of textual gaps, uncertainty, and documentary conventions.