<p>Large language models can generate historical texts with causal organization and narrative coherence, yet their generative process involves no understanding subject in the hermeneutic sense. This phenomenon marks the first occasion on which interpretive structure appears in the absence of understanding proper, raising the question of what epistemic status should be assigned to historical narratives that were once taken to require understanding for their production. The paper further asks what this question reveals about the conditions of historical understanding itself. Drawing on Dilthey, Gadamer, and Ricoeur, the paper distinguishes between understanding proper, referring to the subjective conditions of a historical horizon, belonging to tradition, and epistemic responsibility, and interpretive structure, referring to the structural organization exhibited by output texts, including temporal synthesis, causal connection, and the narrative organization of historical agents’ reasons. On this basis, the paper introduces the concept of Distributed Interpretation to analyze how AI-generated historical narratives acquire interpretive structure through a data layer, a model layer, and a human layer, and to show how the epistemic accountability that was originally unified within a subject becomes dispersed in this process. The paper then advances a weak epistemological position, arguing that such outputs constitute constrained interpretive outputs, an epistemic category between instruments and historical knowledge, which can participate in knowledge production only through the critical intervention of historians. Finally, the paper argues that the CIO case re-illuminates the irreducible functions of subjectivity, historicity, and dialogicality in historical understanding: the core of historical interpretation lies not only in generating narrative structure, but in bringing interpretive results into an epistemic commitment that can be examined, revised, and held responsible.</p>