A layer-resolved diagnostic identifies bias-driven decisions in deep neural networks

Abstract

Modern AI systems can be accurate and confident, but this alone does not reveal whether a decision is well supported by the input. This creates a trust problem because confidence reports how decisive a model is, but not what supports that decisiveness. Here we show that neural-network confidence can be decomposed into input-dependent feature support and input-independent offset support. We formalize this decomposition through the Bias Dominance Index (BDI), a layer-resolved measure quantifying the relative contribution of input-independent offsets to the decision margin, revealing whether confidence is primarily feature-supported or bias-driven. Across convolutional neural networks, a vision transformer and a transformer language model, BDI shows that high confidence can coexist with bias-driven decisions. Layer-resolved analyses map where bias-driven support across network depth. Perturbation analyses further show that the bias component can stabilize performance when readout weights are degraded. Finally, we operationalize decision composition into an acceptance rule that combines confidence and BDI for mechanism-aware auditing and triage. Together, these results position BDI as a general diagnostic of decision composition that distinguishes feature-supported from bias-driven decisions across model families.