Whether to trust: the ML leap of faith

Trust in AI requires understanding the "leap of faith": The paper argues that humans often conflate trust in an AI system with trust in its underlying machine learning technology, but cannot develop genuine trust without understanding what they're actually relying on when they choose to use ML systems.

Traditional explanations of ML reasoning often fail non-experts: The research finds that explaining how a machine learning model works typically doesn't build trust with non-technical users because these explanations are disconnected from users' own mental models and struggle with the inherent complexity of ML systems.

A new approach compares ML decisions to expert human judgment: Rather than explaining ML's reasoning, the paper proposes measuring trust by comparing what an ML system recommends against what a rules-based system (validated by human experts) would recommend—making the differences visible and measurable to users.

The "Leap of Faith matrix" quantifies the gap between human and machine reasoning: By showing users the specific distance between what an ML agent decides and what an expert-validated alternative decides, this approach helps users understand exactly what they're trusting when they adopt an AI system.

Trust should be measured, not just described: The paper challenges the common practice of treating trust as an unmeasurable attitude, instead proposing concrete trust metrics that can be evaluated before users commit to relying on an AI system.