Measuring Error Alignment for Decision-Making Systems

TL;DR

This work introduces two behavioral-alignment metrics, Misclassification Agreement (MA) and Class-Level Error Similarity (CLES), to evaluate how AI and humans err similarly in decision tasks, offering a cheaper complement to Representational Alignment (RA). MA analyzes instance-level joint errors via a misclassification error matrix and Cohen's kappa, while CLES compares error-distribution shapes across classes using a class-weighted Jensen-Shannon divergence. Extensive experiments on synthetic (model-vs-human) and naturalistic datasets show MA provides complementary information to EC and that CLES can proxy MA in data-limited scenarios, with BA metrics correlating meaningfully with RA metrics like CKA. The results suggest BA metrics can inform trustworthiness and value alignment by revealing how closely AI error patterns mirror human expectations, while highlighting domain-dependence and the potential need for multiple metrics to comprehensively assess alignment.

Abstract

Given that AI systems are set to play a pivotal role in future decision-making processes, their trustworthiness and reliability are of critical concern. Due to their scale and complexity, modern AI systems resist direct interpretation, and alternative ways are needed to establish trust in those systems, and determine how well they align with human values. We argue that good measures of the information processing similarities between AI and humans, may be able to achieve these same ends. While Representational alignment (RA) approaches measure similarity between the internal states of two systems, the associated data can be expensive and difficult to collect for human systems. In contrast, Behavioural alignment (BA) comparisons are cheaper and easier, but questions remain as to their sensitivity and reliability. We propose two new behavioural alignment metrics misclassification agreement which measures the similarity between the errors of two systems on the same instances, and class-level error similarity which measures the similarity between the error distributions of two systems. We show that our metrics correlate well with RA metrics, and provide complementary information to another BA metric, within a range of domains, and set the scene for a new approach to value alignment.

Figures (26)

• Figure 1: Different levels of representations. From left to right, it enables the comparison of the decision-making process of two systems at the latent representation level, confidence level, instance level and class level.
• Figure 2: (left) An illustration of dataspace, $\mathcal{X}$, showing decision regions for systems $A$ and $B$ across three classes $1$, $2$ and $3$. Dotted lines indicate decision boundaries for system $A$ (green) and $B$ (red), and decision regions are labelled. The region where system $g$ makes correct classifications, $\mathcal{C}_g$, is shaded appropriately. (right) The same data space, with $3$ data distributions, $D_i$, indicated in black, blue and yellow.
• Figure 3: Example heatmap for MA scores on the Stylized subset from modelvshuman. Darker cells represent a higher value of similarity.
• Figure 4: EC vs MA (left) and MA vs CLES (right) on modelvshuman data, with model-model, model-human and human-human pairs coloured differently, and shaded according to mean accuracy of the pair.
• Figure 5: Spearman's $r$ for each pair of metrics for all system-pairs in both synthetic and naturalistic datasets. Global $r$s measure correlation for all pairs across all datasets; while average $r$s are the mean in-domain $r$-value.