Beyond the Black Box: Equipping Users to Recognize and Challenge AI Bias in Hiring

Abstract

AI is increasingly used in high-stakes decisions; yet, it can reproduce social bias in its outputs. This study tested an end-user-facing, post-processing intervention that pairs model-agnostic counterfactual explanations with brief inoculation training to surface bias at decision time and counter automation bias. Counterfactual inoculation improved bias recognition in hiring, with the strongest effects observed for socially salient cues such as gender and ethnicity. Confidence gains were smaller and inconsistent. Trust did not rise uniformly, indicating calibrated skepticism rather than indiscriminate distrust. Counterfactual audit flags interrupted reflexive acceptance by showing how small input changes affect rankings, while inoculation primed users to look for bias signals, jointly shifting trust toward appropriate calibration. Practical implications include adding simple “what-if” audit flags that work with any model, allowing override or escalation when needed, and providing brief, hands-on training. Limitations include scenario-dependent effects, self-reported outcomes in a simulated workflow, and heteroscedasticity in a subset of models, which was addressed through Welch correction and bootstrapped Analysis of Covariance. Future research should assess durability over time, incorporate behavioral endpoint tools, extend this research to domains beyond hiring, and adapt explanation complexity to the domain and the user. In summary, an inoculation that explains the prevalence of bias in AI hiring tools, plus counterfactuals that expose bias cues offer a practical, interpretable means to make bias visible when decisions are made in and to move end-user trust in AI tools from blind acceptance to calibrated judgment.

Ähnliche Arbeiten