Abstract
Recent advances in generative AI promise to transform economic history by unlocking textual archives, quantifying complex historical concepts, and expanding the empirical toolkit. Yet these opportunities raise fundamental methodological questions. This paper evaluates large language models (LLMs) as tools for digital economic history, comparing their algorithmic foundations with core principles of historical scholarship. We identify three key areas of tension: compromised source criticism due to unclear data provenance, opacity in model training that hinders scholarly review, and probabilistic output that challenges reproducibility. We also diagnose two specific AI-related problems: the “secondary source paradigm,” in which LLMs rely on representations of texts rather than primary sources, and “AI intrusive thoughts,” or the inability of LLMs to respect historical temporality. While these limitations mean AI cannot replace human historical inquiry, we demonstrate practical ways in which it can augment it, such as contextual corpus searches, assistance with digitization workflows, and feedback on methodological framing. Drawing from applications in financial history and correspondence analysis, we propose five guidelines for the responsible and sustainable integration of AI tools into economic history research. The paper is especially relevant to researchers working in developing-country contexts, where archival gaps are common but methodological rigor remains essential. Our contribution bridges technical and historiographical perspectives to support responsible AI adoption in the field.