The Response Shift Paradigm to Quantify Human Trust in AI Recommendations

Trust in AI can be measured by how much humans change their decisions after seeing AI recommendations: The researchers developed a quantitative method that measures "response shift"—the difference between a person's initial answer and their revised answer after receiving an AI recommendation—as a concrete way to assess how much people actually trust AI systems, rather than just asking them subjectively.

Explainability alone doesn't guarantee appropriate trust: The study tested users with AI systems that were either good or poor performers, paired with good, poor, or no explanations. This multi-condition design revealed that the quality of explanations matters, but the findings suggest that simply adding explanations doesn't automatically lead to well-calibrated trust in AI recommendations.

Human cognition is central to responsible AI evaluation: The researchers emphasize that measuring trust and the effectiveness of AI recommendations is fundamentally a problem of understanding how humans think and make decisions, not just a machine learning problem—highlighting that responsible AI requires insights from human psychology and behavior.

Current AI evaluation methods miss the human impact: The paper notes that "the effectiveness of AI recommendations and the trust afforded by end-users are typically not evaluated quantitatively," suggesting that many AI systems in use today lack rigorous measurement of whether they actually help or mislead real people in practice.

This framework enables systematic comparison of explainability methods: By providing a standardized, quantitative way to test different explanation approaches, the paradigm creates a tool for organizations to evaluate which methods for explaining AI actually improve human decision-making in real-world contexts.