Who Determines What Comes Next? Epistemic Initiative in AI-Mediated Art Interpretation

Conversational AI can sustain reflection by asking questions, but every prompt also selects a possible direction for the user. We examine this tension through a field deployment of a voice-based art mediator at the Salar Jung Museum. Across two sequential deployment periods, the system used either a system-led Acknowledge–Interpret–Prompt policy or a later visitor-led policy that responded without routinely appending another question. In 85 system-led and 112 visitor-led sessions, visitor questions constituted 10.7% and 66.4% of conversational turns, respectively, while system-led sessions were substantially longer (5.06 vs. 2.18 minutes on average). Because question opportunities are partly created by the policy itself and the deployments were not randomized, these differences do not establish greater psychological agency. Instead, they expose a design trade-off: system prompting can reduce the burden of deciding what to ask next while also shaping what becomes salient; reduced prompting creates more room for visitor-originated inquiry but can leave some visitors uncertain about how to begin or continue. We distinguish interactional persistence, interpretive participation, and observable epistemic initiative, and propose negotiable initiative: conversational control that shifts according to a user’s demonstrated need for scaffolding rather than remaining continuously with either user or system.