Historical automatic text recognition (ATR) in the Age of LLMs

Historical automatic text recognition (ATR) in the Age of LLMs

Dans le cadre de TranscriboQuest 2026

Conférence (en anglais) de Benjamin KIESSLING (INRIA)

Historical automatic text recognition (ATR) is currently experiencing both a maturation and a paradigm shift. In the hands of humanities scholars, mature tooling, established best practices, and shared research infrastructure have transformed ATR from an experimental technique into a component of everyday scholarly work. At the same time, large language and vision-language models promise excellent recognition results zero-shot, without laborious training-data annotation, training, or fine-tuning. This raises the question whether systematic transcription projects are still needed.

Yet the practical limitations of these systems are increasingly apparent. Hallucination, immense resource consumption, opaque training corpora, and limited reproducibility make them problematic instruments for historical inquiry. More fundamentally, their performance remains uneven: material outside the distributions represented in their pretraining data, including much of the world’s historical writing, is poorly served.

Benjamin Kiessling argues that these developments do not render current aproaches to philologically informed transcription obsolete; they make shared standards all the more urgent. Transcription is not mechanical data entry but an editorial act. Only datasets created according to explicit, communal norms can make those decisions legible, support aggregation across projects, and permit the training and evaluation of specialized, visually grounded methods. The future of historical ATR therefore depends not on model scale alone, but on digital commons that enable informed, reproducible, and well-motivated philological inquiry.