In July 2025, a Microsoft Research team identified historians as the second most AI-exposed profession, with 91 per cent of their typical activities overlapping with current generative AI capabilities. This paper examines what that finding actually measures, where its methodological limits lie, and why the more pressing problem is not AI exposure per se but the structural absence of methodological control over AI-generated outputs. Drawing on the Cambridge response to the Microsoft study and on current research into RLHF-induced sycophancy in large language models, the paper argues that AI systems trained to please rather than to correct pose a specific epistemological risk to a discipline built on interpretive accountability. It then presents VERA-VM as a modular, sequentially enforced framework for AI-assisted art-historical image analysis, with particular attention to its two epistemological reflection modules, ARCHÉ and ICONA.
Andreas Hahn (Sat,) studied this question.