evaluate demographic bias

Existing generative models lack a well-defined fairness evaluation objective when demographic attributes are unspecified, leading to ambiguous assessment criteria. This work systematically formalizes this "missing target" problem and proposes a four-component framework for constructing evaluation targets—comprising the evaluation subject, prior acceptability, allocation mechanism, and operationalization—treating target distribution specification as an integral part of fairness evaluation rather than a preliminary step. Leveraging geographic and occupational priors, the study employs generative auditing and Jensen–Shannon divergence (JSD) to quantify bias, revealing significant deviations between model outputs and geographically derived targets (JSD = 0.508–0.606). Replacing these with uniform-probability targets alters per-cell JSD by 0.279–0.355 on average, underscoring the critical influence of target selection on fairness assessments.

Data uncertainties—such as measurement errors, missing values, and erroneous links—undermine the credibility of policy decisions. Method: This paper proposes a decision-stability-oriented sensitivity analysis framework that shifts the analytical focus from parameter deviation to decision robustness. It introduces an interpretable, decision-level sensitivity metric and integrates counterfactual modeling, hypothesis-driven perturbation sampling, decision boundary tracking, and interactive visualization. Contribution/Results: Evaluated on two real-world policy domains—U.S. presidential vote prediction and childhood lead exposure assessment—the framework significantly enhances policymakers’ awareness of analytical robustness, explicitly delineates credible decision intervals, and provides an actionable confidence assessment tool for data-informed policymaking under data imperfections.

This study addresses the inconsistency among fairness metrics in face recognition, where different measures often yield contradictory conclusions about model bias, thereby exposing the limitations of single-metric evaluation. To tackle this issue, the authors propose the Fairness Disagreement Index (FDI) to quantify the degree of disagreement across multiple fairness criteria and introduce a multidimensional evaluation framework that integrates both error rate disparities and performance-oriented fairness metrics. Through systematic experiments under controlled conditions, they demonstrate that such metric disagreement is pervasive across varying decision thresholds and model configurations, revealing a critical flaw in current fairness assessment practices. The work provides both a novel analytical tool and empirical evidence to support more comprehensive and reliable fairness evaluations in face recognition systems.

This study addresses fairness risks in AI recommendation systems arising from demographic imbalances. Methodologically, it introduces a novel, customizable controlled-dataset construction paradigm; establishes an “affected population vs. overall population” comparative analytical model; designs quantitative bias metrics and a dynamic drift detection mechanism; and integrates these components into the AI auditing platform ITACA. The contributions include: (1) the first operational definition and end-to-end lifecycle monitoring of fairness thresholds across multiple application scenarios; (2) support for training-data calibration, fairness-aware objective formulation, and post-deployment continuous auditing; and (3) real-world validation through Eticas.ai’s auditing practice, delivering actionable fairness guidelines for developers and enabling regulators to formulate verifiable, implementable AI governance policies.

This study investigates how weight and activation quantization affect demographic subgroup bias in large language models (LLMs). We evaluate nine benchmarks spanning stereotypes, toxicity, sentiment, and fairness, using both probabilistic and generative-text-based metrics across diverse model architectures and reasoning capabilities. Results show that quantization generally reduces toxicity and has negligible impact on sentiment, but aggressive compression slightly exacerbates stereotype bias and inter-group unfairness—consistent across models and demographic subgroups. Crucially, this work is the first to uncover the differential mechanisms by which quantization influences distinct bias dimensions. Our findings provide empirical evidence and actionable design insights for reconciling model compression with ethical alignment, highlighting trade-offs between efficiency and fairness that must be explicitly addressed in quantization-aware development pipelines.

This work addresses the absence of a unified and adaptable framework for evaluating gender bias in text-to-image generation models, which hinders alignment with the diverse risk governance requirements across application contexts. To bridge this gap, the authors propose a risk-aligned auditing framework that innovatively introduces the THUMB card mechanism, systematically integrating usage context, manifestations of bias, harm hypotheses, and auditing strategies to enable context-aware bias assessment. Grounded in the European Union AI Act’s risk classification, the framework constructs risk-tiered use-case profiles, a catalog of bias metrics spanning gender prediction, embedding similarity, and downstream task performance, and a contextualized harm typology. This integrated approach yields an interpretable and actionable audit pipeline, substantially enhancing the practicality and relevance of gender bias evaluation in both technical auditing and AI governance.

This study addresses the critical gap in clinical machine learning fairness evaluation by systematically applying an intersectional fairness auditing framework to real-world clinical prediction tasks. Leveraging the All of Us dataset, the authors integrate the FairLogue toolkit, observational fairness metrics, and counterfactual causal analysis to assess model performance across intersecting subgroups defined by race and gender. Their findings reveal substantial performance disparities that remain undetected under conventional single-axis fairness assessments. However, counterfactual experiments demonstrate that most of these disparities persist even after randomizing group identity, indicating that they primarily stem from differences in covariate distributions rather than direct discrimination. These results underscore the necessity and value of intersectional auditing for accurately diagnosing and addressing health inequities in clinical AI systems.

This study investigates the impact of using raw scores versus demographically adjusted scores on classification accuracy and decision fairness in cognitive screening. Through theoretical analysis and empirical validation on the OASIS-3 dataset, it rigorously derives, for the first time, sufficient conditions under which raw scores outperform adjusted scores. The findings demonstrate that common adjustment methods—such as z-score normalization—not only can degrade classification performance under certain conditions but also do not necessarily enhance fairness, thereby challenging the widely held assumption that statistical correction inherently promotes equitable outcomes. This work provides a principled theoretical foundation and practical guidance for score selection in cognitive assessment protocols.

This study addresses inconsistent findings in prior audits of racial and gender biases in large language models (LLMs) during resource allocation. To resolve this, the authors introduce FairFund-Bench, a novel benchmark that systematically reveals how audit formats—such as transparent versus disguised scenarios—significantly influence the direction and magnitude of bias. Grounded in welfare desert theory, they develop a causal need framework and evaluate 14 prominent LLMs across three tasks—scoring, ranking, and allocation—using 600 synthetically generated financial aid requests spanning diverse demographic categories. Results show that while models favor minority groups in individual scoring, they penalize specific demographics in comparative rankings. Bias intensifies markedly under disguised auditing conditions, and the causal framework exerts a far stronger effect on model behavior than demographic attributes alone.