Generative AI as a historical source: source criticism, citation integrity, and the jagged frontier of digital history

Abstract

Between the public release of ChatGPT in late 2022 and 2026, generative artificial intelligence (AI) moved from a computational novelty to a structural feature of historical scholarship, reshaping how primary sources are transcribed, described, analysed, and communicated. This article argues that the decisive methodological shift is not the automation of existing tasks but the arrival of a new kind of object for the historian's craft: the large language model (LLM) itself, which must be read as a historical source rather than trusted as a neutral instrument. Drawing on peer-reviewed evaluations, professional-society guidance, primary legal filings, and documented failure cases, the article develops three connected claims. First, generative models are best understood as an algorithmic cartography of the digitised record whose jagged frontier of competence maps which pasts have been absorbed into training data and which remain silent. Second, the same architecture that enables transcription of damaged manuscripts and large-scale corpus analysis also produces hallucinations and fabricated citations at rates incompatible with the evidentiary standards of the discipline; recent audits and accountability cases in scholarship, government, and the courts illustrate the stakes. Third, the responsible integration of these tools depends on extending traditional source criticism to the model, on non-negotiable verification of every reference, on cryptographic provenance and Indigenous data-governance frameworks, and on assessment redesign rather than prohibition. The article synthesises evidence across document analysis, public history, and pedagogy to propose a programme for a critically literate, symbiotic historical scholarship.