The persuasive power of data visualizations can often lead to issues. In the context of explainable AI (XAI), visualizations can result in over-trust of predictive models. This paper uses a crowdsourced study to demonstrate that providing accurate (but superfluous or irrelevant) data in model explanations can lead to unjustified trust and other positive beliefs about a model, even when it is clearly discriminatory and unfair.
Our findings suggest that XAI designers and developers must consider the implicit or explicit rhetoric of their work and be wary of the potential for visualizations to imbue models with unearned trust.
Blogger's Review: This study highlights the potential risks of data visualization in the field of explainable AI, emphasizing the need for designers to be cautious in model explanations to avoid misleading users about the trustworthiness of models, especially in fairness-related contexts.