Tracing Communicative Motive in LLM-Generated Visualizations: An Activity-Theory-Informed Artifact Analysis

Abstract

LLM-based visualization workflows can translate broad prompts into data transformations, analysis code, and visual encodings [2,5], but this automation makes human-specified communicative motives difficult to trace. We examine this traceability problem through a work-in-progress artifact analysis of an Activity-Theory-grounded prompting workflow [4]. Across three public datasets and three intended VisAct motives—Inform, Provoke, and Discover—we generated 135 Vega-Lite visualizations using three frontier LLMs and coded 269 visual components [3]. Each component was analyzed by composition strategy, insight need, visual form, claim, rhetoric with codes adapted from visualization task typologies, the nested model, VisActs, and visualization rhetoric [1, 3, 6, 7]. The resulting patterns suggest that motive conditions are associated with different artifact-level design pathways and recurring misalignments among motive, insight need, form, and framing. We contribute an exploratory coding workflow for studying communicative motive realization in LLM-generated visualizations.