Automating the Path: An R&D Agenda for Human-Centered AI and Visualization

TL;DR

The paper addresses how visualization and visual analytics can robustly integrate human-centered AI amid rapid advances in generative AI and foundation models. It proposes a design space that maps four HCAI capabilities—amplify, augment, empower, enhance—onto the four visual sensemaking phases—prepare, explore, schematize, report—to guide an R&D agenda. Through reviews of existing tools (e.g., Wrangler, DataSite, grounded theory helper, DG Comics) and discussions of ethical considerations, it outlines concrete opportunities, pitfalls, and governance needs across all phase-cells. The authors argue for a convergent future where AI augments human sensemaking without compromising agency, transparency, or accountability, promoting open, modular tools and human-centric interfaces. Overall, the framework aims to accelerate responsible AI-enabled visualization, enabling both domain experts and researchers to discover the unexpected while maintaining rigorous interpretability and control.

Abstract

The emergence of generative AI, large language models (LLMs), and foundation models is fundamentally reshaping computer science, and visualization and visual analytics are no exception. We present a systematic framework for understanding how human-centered AI (HCAI) can transform the visualization discipline. Our framework maps four key HCAI tool capabilities -- amplify, augment, empower, and enhance -- onto the four phases of visual sensemaking: view, explore, schematize, and report. For each combination, we review existing tools, envision future possibilities, identify challenges and pitfalls, and examine ethical considerations. This design space can serve as an R\&D agenda for both visualization researchers and practitioners to integrate AI into their work as well as understanding how visualization can support HCAI research.

Figures (4)

• Figure 1: Stanford Wrangler. The left panel presents a transformation history tracking data provenance, a transform selection menu for manual ETL operations, and contextually suggested transformations based on current data selection. Interactive parameters within transform descriptions (shown in bold) enable direct manipulation of the ETL pipeline. The right panel displays the resulting data table with integrated data quality meters above each column, providing immediate feedback on transformation outcomes.
• Figure 2: DataSite. This tool enhances exploratory data analysis through a human-AI collaborative framework. As analysts interactively develop and refine hypotheses through visual exploration, the system continuously surfaces analytically significant features in the Feed View (right), similar to a social media stream. This architecture maintains the analyst's agency in visual discovery while augmenting their analytical reasoning capabilities with AI-driven computational support, enabling more fluid and comprehensive exploration of data patterns, relationships, and anomalies.
• Figure 3: Grounded theory helper. The interface supports human-AI collaborative grounded theory analysis through computational linguistics. Server-processed text metadata (parts of speech, entities, information content) powers coordinated visualizations that facilitate the schematization of qualitative data. These interactive views help analysts identify concepts and relationships while building coherent analytical structures during the iterative coding process, blending computational support with human interpretive authority.
• Figure 4: DG Comics. The tool employs a human-AI collaborative approach to dynamic graph communication, featuring a summary view, interactive filtering sliders, a graph comic view, character tables, and a timeline. Users can access additional perspectives through a node attribute table or a community view. The system supports effective knowledge dissemination through mental map preservation techniques and visual community change representations, enabling stakeholders to explore actionable insights and make informed decisions.