The Human-Data-Model Interaction Canvas for Visual Analytics

Visual analytics requires intentional design of human-data-model interactions: The paper identifies that effective data-driven decision-making depends on systematically characterizing how humans, data, and models work together—not treating them as separate components. This suggests responsible AI systems should be designed with explicit attention to how these three elements interact rather than optimizing models in isolation.

Human knowledge and feedback loops are essential, not optional: The HDMI Canvas emphasizes "human knowledge externalization and forms of feedback loops" as core to visual analytics processes. For responsible AI, this means building systems that actively capture human expertise and enable people to correct or refine model outputs, rather than treating human input as a one-time training phase.

Explainability and interpretability should be standard model contributions: The paper highlights that "interpretable and explainable AI" are important model contributions beyond conventional outputs. This supports the principle that AI systems used in decision-making should be designed to show their reasoning, not just deliver predictions, enabling stakeholders to understand and question model decisions.

Diverse stakeholder involvement improves AI system design: The canvas is "optimized for external stakeholders, improving VA outreach, interdisciplinary collaboration, and user-centered design." This suggests that responsible AI development benefits from involving people with different expertise and perspectives early in design, rather than having technical teams work in isolation.

Process frameworks help differentiate between different types of AI contributions: By providing a structured way to characterize "diverse roles of humans, data, and models," the canvas enables clearer communication about what each actor contributes, which can help non-technical stakeholders better understand and evaluate AI systems in their contexts.