Based on my opening keynote at DH2025, this article examines the impact that transformer-based machine learning brings to the interpretive work of historians. The discussion begins by revisiting an earlier phase of my research, which focused on digitally mediated, multiscale analysis – or what I refer to as ‘digital (re)reading’ – using database queries to detect structural patterns in digitized historical sources. Building on this foundation, I turn to the present, where my team and I investigate how large language models (LLMs) and vision–language models (VLMs) can support algorithmic reading across heterogeneous and semantically complex corpora. This new phase of inquiry explores how such models enable semantic, stylistic, sentiment and multimodal analysis, moving decisively beyond the constraints of keyword search and frequentist approaches. The article also introduces the DeepPast initiative, a modular artificial intelligence (AI) framework designed to promote the use of pluggable, task-specific components that run on low-power hardware rather than hyperscale, monolithic systems. The DeepPast architecture supports multiple interpretive modes, allowing historians to engage in a structured and purposeful dialogue with an AI assistant that functions as a genuine research partner. The discussion concludes with reflections on how to keep historians in the loop and to produce AI-assisted historical research marked by greater interpretive nuance and sophistication.
Javier Cha (2026) studied this question.