Artificial intelligence systems are widely evaluated based on predictive accuracy, confidence calibration, and output performance. However, increasing evidence suggests that these metrics are insufficient to explain real-world decision failures. Systems can generate highly accurate predictions while simultaneously failing to recognize, interpret, or act upon critical signals embedded within complex environments. This paper introduces Signal Sensitivity as a foundational concept for analyzing and engineering decision systems. While prior research - such as Signal Detection Theory and metacognitive sensitivity - focuses on the ability to distinguish signal from noise at the level of isolated judgments, these approaches do not account for how signals propagate through decision architectures. We define Signal Sensitivity as a system-level property: the capacity of a decision system to detect, interpret, integrate, and respond to relevant signals across its full operational pipeline. This perspective shifts the focus from detection accuracy toward structural responsiveness, highlighting how signals may be lost, distorted, or ignored at different stages of decision-making processes. The paper develops a formal framework for Signal Sensitivity within the context of Decision Engineering Science™ (DES), positioning it as a critical layer between prediction and decision execution. We introduce key measurable dimensions, including Signal Detection Rate (SDR) and Missed Signal Rate (MSR), and examine how these metrics relate to decision quality and system performance. By extending beyond traditional detection-based models, this work contributes a novel lens for understanding why decision systems fail - even when predictive models perform well. The framework has direct implications for artificial intelligence, human–AI collaboration, governance systems, and high-stakes decision environments such as finance, healthcare, and policy. This paper positions Signal Sensitivity as an essential component of next-generation decision architectures, enabling more robust, adaptive, and context-aware systems.
Aleksandra Pinar (Fri,) studied this question.