Atlas of Human-AI Interaction

• Atlas of Human-AI Interaction is a comprehensive, interactive knowledge graph that maps empirical findings, research themes, and causal mechanisms of human–AI relationships.
• It employs advanced LLM-based extraction and unsupervised machine learning to cluster over 1,000 peer-reviewed studies, offering detailed insights and thematic clarity.
• The resource enables actionable research by visualizing quantitative measures, causal pathways, and evolving interaction patterns to guide design, policy, and further meta-analysis.

The Atlas of Human-AI Interaction is presented as a comprehensive, systematized mapping of empirical findings, research themes, and causal mechanisms underlying the diverse relationships between humans and AI systems. This resource is instantiated as an interactive knowledge graph, constructed using LLM-based knowledge extraction from over a thousand peer-reviewed HCI papers and community data from real-world online platforms. The atlas is both an analytical and practical tool, enabling researchers and practitioners to navigate, synthesize, and interrogate the empirical landscape of human–AI interaction with an unprecedented level of granularity and causal clarity (Pataranutaporn et al., 14 Sep 2025).

1. Interactive Knowledge Graph and Web Interface

The core of the atlas is an interactive web-based dashboard designed for exploration of large-scale datasets aggregating human–AI interactions. The interface enables users to:

• Pan, zoom, and filter across semantically clustered conversation clouds—clusters reflect varying interaction modes, from utilitarian and technical, to deeply social or emotional (e.g., AI companionship, troubleshooting, artistic co-creation).
• Access both macroscopic thematic distributions and detailed drill-downs showing representative posts, empirical findings, or statistics for each cluster.
• Visualize quantitative measures and causal linkages in a navigable 2D or 3D embedding space, connecting abstract empirical regularities to rich, contextual examples extracted from actual user experiences or research studies.

This implementation facilitates discovery-oriented navigation, supports systematic reviews across fragmented literatures, and allows for contextualized, causal analysis of interaction outcomes.

2. Data Extraction Methodology and Thematic Clustering

The atlas is powered by an empirical pipeline that combines unsupervised machine learning and LLM-based interpretation:

• Data Aggregation: The system collects large corpora of user-generated posts from Reddit communities focusing on AI companionship, along with metadata associated with user behaviors.
• Semantic Embedding: Textual data is mapped to high-dimensional embedding spaces (e.g., Qwen3 Embedding, 768 dimensions). Dimensionality reduction using UMAP projects this onto a 2D space suitable for visualization.
• Cluster Determination: The optimal number of thematic clusters is chosen using the elbow method, computed via analysis of Within-Cluster Sum-of-Squares (WCSS), rendered in LaTeX as:

$$\text{WCSS}(K) = \sum_{i=1}^{K} \sum_{x \in C_i} \|x - \mu_i\|^2$$

where $C_i$ is the set of points in cluster $i$ and $\mu_i$ its centroid.

• Cluster Interpretation: Each resultant cluster is auto-labeled and summarized using LLMs (Claude Sonnet 4, GPT-5 variants), with further human researcher validation to ensure labeling fidelity and thematic consistency.

This pipeline yields a granular, interpretable map capturing the multi-dimensionality of human–AI relationships, with each cluster representing a distinct motif or theme (e.g., emotional intimacy, breakdowns from technical changes, social support).

3. Empirical Findings and Causal Relationships

Moving beyond mere topic clustering, the atlas computes and visualizes cause-and-effect relationships across empirical findings:

• Extraction identified 2,037 distinct empirical findings, many mapping direct or mediated causal links between interaction design decisions and user outcomes.
• For instance, the system reveals that emotional bonds with AI emerge disproportionately among users who initially encountered AI in utilitarian contexts (e.g., productivity tools), compared to those actively seeking companions.
• It identifies effects of disruptive model updates as triggers for grief responses and relationship breakdowns—a pattern analogous to real human attachment and loss.
• Analysis of platform usage reveals significantly higher emotional anthropomorphization and dependency amongst users of general-purpose conversational agents versus dedicated AI companion bots.

These findings are not simply correlations, but are derived using automated LLM-based, evidence-driven causal inference, with interface affordances to drill into quantitative metrics or representative qualitative evidence.

4. Implications for Research, Design, and Meta-Analysis

The atlas functions as both a scientific instrument and a methodological blueprint:

• Research Gap Discovery: The interactive clustering and causal mapping enables users to isolate underexplored or disconnected areas, supporting targeted literature review and hypothesis generation.
• Design Interventions: By mapping causal pathways (e.g., from inadvertent interaction to emotional dependency), the atlas enables the anticipation and mitigation of risks (such as the need for continuity controls to protect users from abrupt system changes).
• Computational Synthesis: The technical framework, leveraging LLM interpretability and embedding techniques, demonstrates a scalable new modality for computational meta-science—not just aggregating similarity, but quantifying impact pathways.
• Temporal Analysis: The system supports real-time monitoring of evolving trends, offering a benchmark for policy makers and system designers as models and communities evolve.
• Policy and Governance: By surfacing quantitative links and evidence, the atlas aids in evidence-based policy development, regulation, and ethical design by providing granular views of sociotechnical risks and benefits.

5. Methodological Highlights: Embedding, Clustering, and Interpretation

The system’s methodology synthesizes mathematical and interpretive layers:

| Component | Approach/ Representation | Scholarly Context | |---------------|----------------------------------------------------------|-------------------------------------------------| | Data | Reddit AI communities, empirical findings from literature | Massive scale for broad representativeness | | Embedding | Qwen3 768-dim semantic vectors, UMAP to 2D | Preserves semantic and contextual distinctions | | Clustering | Elbow method on WCSS, followed by LLM-based theme labeling| Automatic, scalable, interpretable clustering | | Interpretation| LLM-driven summaries, human validation | Ensures thematic fidelity and transparency |

This multi-stage process directly connects mathematical abstraction (embedding, clustering) to interpretive synthesis (LLM plus expert curation), providing robustness and scalability for ongoing empirical synthesis.

6. Application for Gap Analysis and Future Directions

Researchers and practitioners can use the atlas to:

• Drill down into specific behaviors (e.g., grief after technical update) and trace their empirical roots and consequences.
• Search across the hierarchical taxonomy to isolate novel clusters or phenomena for investigation (e.g., unstudied interaction modes with visual sharing).
• Employ visual analytics combined with mediation models (e.g., Baron and Kenny’s model, bootstrapped intervals) for hypothesis testing and longitudinal studies.
• Monitor real-time topic evolution as AI systems, communities, and user needs change.

A plausible implication is that this computational approach to literature synthesis—transcending static keyword clusters—may become the new norm for aligning research, design, and policy in the field of human–AI interaction.

7. Significance for Human–AI Interaction Science

The atlas marks a paradigm shift by:

• Systematizing the causal landscape of human–AI interaction across the spectrum from utilitarian, artistic, to deeply social and emotional modalities.
• Demonstrating that empirical synthesis can be enhanced and, in part, automated via LLMs, semantic embedding, and evidence mapping.
• Supporting the entire research–design–policy pipeline, from hypothesis discovery through to the specification of actionable design or regulatory guidelines.
• Providing a practical, interactive tool for the HCI and AI research communities to bridge gaps, surface emergent risks, and iteratively refine their understanding of an increasingly complex socio-technical domain.

This resource redefines the benchmark for evidence-based design and meta-analysis in the fast-evolving area of human–AI interaction, and suggests a broader trajectory for computational meta-science in allied fields (Pataranutaporn et al., 14 Sep 2025).