Project Overview
The Atlas of Human-AI Interaction is an innovative framework that maps the complex landscape of empirical findings in human-AI interaction research. Unlike traditional literature reviews that organize papers by themes or topics, this project traces connections between studies based on cause-effect relationships in their empirical findings.
By analyzing over 1,000 papers from major HCI venues, our team has used large language models (LLMs) to extract and connect research findings as structured triplets in the form [cause, relationship, effect]. The resulting knowledge graph reveals patterns of AI's influence on human experience across different contexts, identifies structural holes in the research landscape, and highlights promising areas for future investigation.
Our interactive visualization platform allows researchers and practitioners to explore the landscape of human-AI interaction findings through multiple views: a complete network visualization, focused cluster views, and detailed node information displays. The platform enables users to identify key patterns in how different AI technologies affect human experiences, discover well-established relationships, and pinpoint gaps where further research is needed.
The atlas serves both as a navigational aid for understanding the current state of knowledge and as a foundation for developing more nuanced, context-aware AI systems that better support human needs and capabilities. Our methodology also demonstrates the potential of LLMs as research synthesis tools for processing and connecting findings across large bodies of literature.
This project contributes to the field by providing an evidence-based understanding of human-AI interaction patterns, revealing both areas of convergence (such as transparency mechanisms and trust development) and significant gaps (such as longitudinal effects and reciprocal adaptation). As AI systems become increasingly integrated into daily life, this systematic mapping of empirical knowledge offers valuable insights for both research priorities and evidence-informed design decisions.
Methodology
Our methodology employs a systematic approach to extracting, processing, and visualizing research findings from academic literature through five main stages:
- Research Abstract Collection: We gathered over 1,300 research abstracts related to human-AI interaction from major academic databases including ACM Digital Library, IEEE Xplore, Springer Nature Link, and ArXiv.
- Findings Triplet Extraction: Using large language models (LLMs), we extracted structured triplets in the form [cause, relationship, effect] from each paper's findings. Each triplet classifies elements into human, AI, or concept/object categories, with relationships coded as INCREASES, DECREASES, or INFLUENCES.
- Triplet Normalization: We processed the extracted triplets to standardize terminology, merge synonymous concepts, and prepare them for network analysis.
- Semantic Clustering: We applied advanced clustering techniques to organize the concepts into meaningful groups, identifying 6 human-related clusters, 12 AI technology clusters, and 8 concept/object clusters.
- Graph Transformation: The processed data was converted into a knowledge graph representation, enabling interactive visualization and analysis of the research landscape.
Key Findings
Common Themes
Our analysis revealed several dominant patterns across the research landscape. AI output characteristics emerged as the most frequently studied causal factor, followed by generative AI systems and chatbots. On the effect side, students and general participants were the most common subjects, highlighting the prevalence of educational applications and controlled experimental settings in HAI research.
The most prominent relationship patterns included connections between AI systems and educational contexts, explanation mechanisms and trust development, and design characteristics influencing design practitioners themselves. This reveals the field's current emphasis on transparency, education, and design processes.
Research Clusters
The knowledge graph organization identified distinct clusters across human, AI, and conceptual domains. Human clusters ranged from artistic individuals to healthcare professionals and knowledge workers. AI technology clusters separated different paradigms including natural language processing, autonomous systems, conversational agents, and domain-specific applications. Concept clusters distinguished between concrete tools, disciplinary domains, and abstract dimensions like ethics and perception.
Underdeveloped Areas
Our structural analysis identified significant gaps in the research landscape, representing opportunities for future investigation. These include:
- Integrated AI output design frameworks that connect generation with human interpretation
- Cross-domain collaboration models integrating collaborative theories with AI system design
- Domain-adaptive explainability frameworks tailored to specific contexts
- Better integration of large language models with traditional HAI paradigms
- More comprehensive educational AI integration frameworks
Interactive Tools
The project includes an interactive web-based visualizer that displays the knowledge graph as a 3D network. Built using Svelte.js, Three.js, and D3.js, this tool allows researchers and practitioners to explore connections between findings, focus on specific clusters, and examine detailed information about individual nodes and their relationships. Multiple views provide different perspectives on the data, from the complete network to focused clusters and individual node connections.
Future Directions
Several promising directions for future work could enhance the atlas's utility and scope. These include:
- Continuous data addition and updates to maintain relevance
- Extended relationship types beyond the current causal framework
- Refined extraction processes with more sophisticated LLM approaches
- More intensive connectivity analysis to uncover hidden patterns
As AI systems become increasingly integrated into daily life, maintaining a clear understanding of their effects on human experience grows in importance. This atlas represents one approach to supporting more informed development and deployment of these technologies through systematic knowledge mapping.