City Research Online ***TEST***

Abstract

Large Language Models (LLMs) are increasingly used for generating and adapting visualizations for different user groups. While recent efforts have focused on adapting visualizations to users’ cognitive and perceptual abilities, how LLMs cater to the distinct interests and subjective priorities of various stakeholder groups remains largely unexplored. Specifically, LLMs utilise rhetorical elements to prioritise data stories, which can shape user interpretation. We present a systematic approach to assessing how LLMs adapt their visualization rhetoric to match the priorities of different user personas in a healthcare context. Based on qualitative interviews with population health stakeholders, we demonstrate LLMs’ capabilities for (i) understanding user tasks and priorities from interview data, (ii) adapting visualizations to these priorities, and (iii) justifying design choices for the adaptations. Population health data presents an excellent space for experimentation, given: (a) the diversity of stakeholders (e.g., commissioners, population health experts, data analysts, and the public); and (b) the varied purposes and key messages for which visualizations are designed. We reflect on patterns in LLM reasoning about persona-specific design choices—in light of an established analytical framework for rhetorical visualization—and propose open questions to promote safer, more responsible practices in LLM-assisted visualization.

Publication Type:	Conference or Workshop Item (Paper)
Publisher Keywords:	Generative AI, Visualization, User Tasks
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Science & Technology School of Science & Technology > Department of Computer Science
SWORD Depositor:	Symplectic Administrator

Export

Downloads