Academic References

Examines how technologies reproduce racial hierarchies and proposes frameworks for understanding algorithmic discrimination as a continuation of historical patterns.

Foundational work on how classification systems embed social and political assumptions, essential for understanding bias in feature engineering.

Comprehensive examination of how AI systems encode power structures and historical patterns of discrimination.

Applies feminist theory to data science, emphasizing how power structures shape data collection, analysis, and representation.

Case studies demonstrating how algorithmic systems intensify surveillance and control of economically marginalized communities.

Analysis of how search algorithms perpetuate and amplify racist and sexist stereotypes through their ranking mechanisms.

Accessible introduction to algorithmic harms across domains including education, employment, and criminal justice.

Examines the opacity of algorithmic systems and advocates for transparency and accountability in automated decision-making.

Landmark legal analysis of how big data analytics can produce discriminatory outcomes even without discriminatory intent, connecting data mining to disparate impact doctrine.

Comprehensive textbook covering the mathematical foundations, societal context, and practical limitations of algorithmic fairness.

Proves impossibility of simultaneously achieving calibration and equal error rates across groups when base rates differ.

Comprehensive survey of fairness definitions, their relationships, and limitations in practical applications.

Introduces individual fairness as treating similar individuals similarly, foundational for metric-based fairness approaches.

Defines equalized odds and equal opportunity metrics, establishing key group fairness constraints.

Fundamental impossibility theorem showing calibration and balance constraints cannot be simultaneously satisfied.

Comprehensive taxonomy of bias types, fairness definitions, and mitigation techniques in machine learning.

Accessible overview of major fairness definitions with clear explanations and relationships.

Landmark study demonstrating intersectional accuracy disparities in facial recognition, revealing up to 34% error rate differences.

Introduces methods to detect hidden disparities at subgroup intersections that aggregate metrics miss.

Reveals how using healthcare costs as a proxy for health needs systematically disadvantaged Black patients.

Documents how historical bias in policing data propagates through predictive systems, creating feedback loops.

Comprehensive taxonomy of bias sources across the ML lifecycle from data collection to deployment.

Proposes methods for measuring and communicating fairness in ranking systems.

Synthesizes AI ethics principles into a unified framework: beneficence, non-maleficence, autonomy, justice, and explicability.

Critical analysis of principle-based AI ethics, emphasizing the need for virtue ethics and institutional practices.

Examines how fairness conceptions differ across cultural contexts, with implications for global AI deployment.

Argues for democratic participation in defining fairness criteria, connecting technical and political dimensions.

Explores non-Western ethical traditions including Ubuntu, Confucian ethics, and Indigenous value systems for AI fairness.

Comprehensive analysis of 84 AI ethics guidelines, identifying convergence and divergence in principles.

Foundational work on intersectional analysis and the matrix of domination, essential for understanding compound discrimination.

Introduces the concept of intersectionality, demonstrating how single-axis frameworks fail to capture compound discrimination.

Documents systematic gender bias in data collection and technology design across domains.

Argues for applying critical race theory to algorithmic fairness research, questioning how race is operationalized in ML.

Examines how racial categories are constructed and reinforced through scientific and technological practices.

Investigative analysis of COMPAS recidivism algorithm, catalyzing public debate on algorithmic fairness.

Classic study documenting racial discrimination in hiring, foundational for understanding bias in employment algorithms.

Comprehensive examination of predictive policing technologies and their implications for racial justice.

Historical analysis of actuarial methods in criminal justice, demonstrating longstanding patterns of discrimination.

Historical context for understanding racial disparities in criminal justice data used to train ML systems.

Examines how automatic gender recognition systems harm transgender individuals through binary classification.

The world's first comprehensive AI regulation. Art. 5 defines prohibited practices; Annex III lists high-risk AI use cases requiring conformity assessment.

Comprehensive review challenging the scientific validity of automated emotion recognition — foundational for the EU AI Act's prohibition of emotion recognition in workplace/education (Art. 5(1)(f)).

Critical legal analysis of the AI Act's risk-based approach, including the definition of high-risk AI systems.

Regulatory guidance for AI-based credit scoring under the EU AI Act's Annex III, Area 5(b) high-risk classification.

Right not to be subject to solely automated decisions with legal effects. Key for credit scoring and employment decisions.

Foundational EU anti-discrimination directive applying to employment, education, social protection, and access to goods and services.

Algorithm targeted dual-nationality families for welfare fraud, causing financial ruin for ~26,000 families. Led to government resignation.

SyRI welfare fraud detection system ruled a violation of ECHR Art. 8 (right to privacy). First European court ruling striking down an algorithmic welfare system.

CJEU ruled that credit scoring by SCHUFA constitutes automated decision-making under GDPR Art. 22, requiring transparency and the right to explanation.

Austrian AMS employment scoring algorithm penalised women (due to care responsibilities) and disabled jobseekers, allocating fewer resources.

Investigation showing 78% of Gangs Matrix entries were Black individuals despite comprising 13% of the borough population.

UK exam grading algorithm systematically disadvantaged students at state schools and from ethnic minority backgrounds compared to private schools.

Comprehensive EU-wide survey on discrimination experienced by minorities, including Roma, immigrants, and religious minorities.

Introduces the exponentiated gradient reductions approach, converting fair classification into a sequence of cost-sensitive classification problems.

Proposes convex relaxations for incorporating fairness constraints directly into classifier training objectives.

Develops proxy-Lagrangian methods for optimizing non-differentiable fairness constraints during model training.

Introduces the rejection option classifier for post-processing fairness, modifying predictions only in the uncertainty region near the decision boundary.

Foundational work on differential privacy, providing mathematical guarantees for protecting individual data in statistical analyses.

Develops methods for estimating heterogeneous treatment effects, enabling subgroup-level causal analysis in fairness experimentation.

Seminal framework for mediation and moderation analysis, foundational for understanding proxy discrimination pathways.

Introduces the Sequential Probability Ratio Test (SPRT), the statistical foundation for early stopping in fairness A/B testing.

Landmark ruling finding ballot-box naturalisations discriminatory. Led to 2018 BüG revision requiring written decisions with reasoning.

Documents how Swiss permit types (Ausweis B, C, F, N) function as nationality proxies in employment, housing, and credit decisions.

Field experiments showing 20-50% fewer callbacks for applicants with Balkan, Turkish, and African names compared to Swiss-German names in housing.

Monitors discrimination across employment, housing, education, and public services. Key source for Swiss-specific bias patterns.

Sozialhilfe receipt can trigger permit consequences under AIG Art. 63, creating a punitive feedback loop disproportionately affecting immigrants.

Switzerland's modernised data protection law, aligned with GDPR principles. Relevant for automated profiling and decision-making in Swiss AI systems.

Evaluations and studies on RAV/ORP employment profiling systems, including findings of bias against women and older workers.