Interactive Model Cards: A Human-Centered Approach to Model Documentation

Interactive model cards significantly improve comprehension for non-expert users: The study found that augmenting static model cards with interactive features and exploration affordances helps analysts without formal ML/NLP training better understand model documentation, addressing a critical gap between how models are currently documented and who actually deploys them.

Standard model cards fail to support appropriate model use: The research identified that traditional static model cards are "tailored primarily to individuals with ML or NLP expertise," meaning non-expert analysts lack the guidance needed to deploy models responsibly and recognize when models are unsuitable for their intended applications.

Design choices matter for building appropriate skepticism: Among the six conceptual dimensions identified (stakeholders, design, guidance, understandability & interpretability, sensemaking & skepticism, and trust & safety), the paper emphasizes that thoughtful design of interactive elements is essential for helping non-experts develop healthy skepticism about model limitations rather than over-trusting AI systems.

Context and organizational factors are critical but often overlooked: The findings highlight that effective model documentation must consider "broader sociotechnical contexts and organizational" factors—suggesting that interactive cards alone are insufficient without attention to how models are actually used within real workplace environments and decision-making processes.

Stakeholder involvement in design is essential: The study's two-phase approach (first consulting ML/NLP/ethics experts, then testing with actual non-expert users) demonstrates that responsible AI documentation requires input from diverse stakeholders, not just technical experts designing for other experts.