Reading the Unreadable: Artificial Intelligence for Text Interpretation in Degraded Manuscripts

This study addresses the methodological and epistemological implications of employing artificial intelligence in the study of severely damaged manuscripts, proposing a paradigm shift from traditional Handwritten Text Recognition (HTR) systems that fail under regimes of severe material degradation. Three reference computer vision architectures are compared on the multispectral MADAM dataset: object detection (Faster R-CNN), semantic segmentation (U-Net), and instance segmentation (Mask R-CNN), evaluated on incremental training subsets ranging from 6 to 84 patches. Semantic segmentation outperforms both object detection and instance segmentation across most metrics, achieving stable performance from 24 annotated patches; conversely, Faster R-CNN and Mask R-CNN architectures require larger datasets to avoid divergence during training (Carcagnì et al. 2026). However, U-Net encounters a structural limitation in case of thin or nearly transparent strokes where residual ink is insufficient to support segmentation; under such conditions, textual completion requires the expert’s judgment. This work, therefore, justifies a human-in-the-loop (HITL) approach in which algorithmic analytical capabilities and philological, linguistic, and semiotic expertise operate in a complementary manner throughout the entire recognition pipeline. A new hermeneutic paradigm, where the acronym AI alludes to a necessary collaboration between Artificial Intelligence and Artisan Intelligence –for the safeguarding of Artistic Intelligence of humankind.