Abstract: This article examines the use of large language models in reading nineteenth-century diplomatic archives,with particular reference to the British Foreign Office FO17 series. It argues that large language models can substantially lower the threshold for working with handwritten archival materials,but that their generative logic also introduces new risks for historical interpretation. By relying on contextual associations and learned linguistic patterns,these models often produce coherent and readable transcriptions at the cost of smoothing over the material,irregular,and historically specific features of the original documents. The article describes this tendency as “fluency-induced occlusion”:a process through which uncertain,fragmented,or anomalous textual signals are absorbed into a seemingly fluent and plausible output. Such occlusion may generate systematic hallucinations and,when combined with the uneven distribution of training data across languages,may reinforce a form of linguistic centralism in AI-assisted research. The article therefore argues that historians should not treat AI merely as a tool for efficiency,nor should they rely solely on “defensive proofreading. ” Instead,algorithmic bias itself can be used as a reverse lens through which to examine the logic,materiality,and interpretive tensions of historical documents. In this sense,human-AI collaboration invites historians to rethink the agency,responsibilities,and methodological boundaries of historical research in the digital age.