This thesis investigates gendered algorithmic discrimination in the labor context by combining a doctrinal analysis of European and Italian law with technical-empirical evaluation of the mechanisms and effects of bias in AI systems used in human resources. The first section reconstructs the historical and legal development of the prohibition against sex discrimination in European law from the Treaty of Rome through subsequent Treaty provisions and fundamental rights instruments and maps the current regulatory framework, with particular attention to Directive 2006/54, the recent transparency and pay-related instruments of 2023–2024, and the implications of the AI regulatory framework. The second section introduces types of bias cognitive, gendered, and algorithmic and explains how socio-technical processes translate human prejudices into modelled outcomes. The third section traces how algorithmic discrimination emerges across the ML lifecycle (data poisoning, biased labeling, and black-box opacity) and examines concrete impacts across recruitment, promotion, performance evaluation, and workplace monitoring. Four case studies (Amazon, Facebook, Apple Card, and HireVue) connect theoretical mechanisms to empirical instances of disparate outcomes and opacity-related harms. The final section assesses legal and policy strategies for mitigation, evaluating EU-level responses such as risk-based AI regulation, pay-transparency measures, and shifts in burden of proof, alongside recent Italian initiatives addressing AI governance, HR transparency, and gender-equality certification. Drawing on the doctrinal and empirical findings, the thesis proposes operational recommendations for policymakers and HR practitioners (continuous algorithmic auditing, context-sensitive fairness metrics, datasheets/model cards, gender impact assessments, and participatory design) and outlines research priorities including context-aware metrics and intersectional monitoring.
La tesi esamina la discriminazione algoritmica di genere nel contesto del diritto del lavoro, combinando un'analisi normativa europea e italiana con una valutazione tecnico-empirica delle cause e degli effetti dei bias nei sistemi di intelligenza artificiale applicati alle pratiche HR. Partendo dall’evoluzione storica del divieto di discriminazione di genere nell’ordinamento europeo, dal Trattato di Roma alle disposizioni attuali del TFUE, dalla Carta di Nizza alla Convenzione europea dei diritti umani, la prima parte ricostruisce il quadro giuridico e le direttive rilevanti, con particolare attenzione alla Direttiva 2006/54, alle recenti Dir. 2023/970 e 2024/1500 e al Regolamento AI Act (Reg. 2024/1689). La seconda parte introduce le tipologie di bias cognitivi, di genere e algoritmici e spiega i meccanismi tecnici attraverso i quali l’AI può riprodurre disuguaglianze. La terza parte analizza come la discriminazione algoritmica si genera nelle diverse fasi del ciclo di vita dei sistemi (data poisoning, labeling, opacità delle black box) e illustra impatti concreti nelle pratiche di reclutamento, gestione, promozione e valutazione delle performance; include casi studio su Amazon, Facebook, Apple Card e HireVue per collegare teoria e prassi. Nell’ultima parte la tesi valuta strategie normative e operative per la mitigazione: l’inquadramento dell’AI Act sull’“alto rischio”, la Direttiva sulla trasparenza salariale, l’inversione dell’onere della prova e le politiche di governance aziendale; a livello nazionale si considerano la legge italiana sull’IA (l.132/2025), il decreto sulla trasparenza e gli strumenti di certificazione e monitoraggio della parità di genere. Sulla base dell’analisi, si propongono raccomandazioni pratiche per policy-maker e operatori HR (auditing algoritmico continuo, datasheets/model cards, procedure di valutazione d’impatto di genere, misure partecipative) e indicazioni di ricerca future su metriche contestuali e intersezionalità.
La discriminazione algoritmica di genere sul luogo di lavoro
CECCHIN, GIOIA
2025/2026
Abstract
This thesis investigates gendered algorithmic discrimination in the labor context by combining a doctrinal analysis of European and Italian law with technical-empirical evaluation of the mechanisms and effects of bias in AI systems used in human resources. The first section reconstructs the historical and legal development of the prohibition against sex discrimination in European law from the Treaty of Rome through subsequent Treaty provisions and fundamental rights instruments and maps the current regulatory framework, with particular attention to Directive 2006/54, the recent transparency and pay-related instruments of 2023–2024, and the implications of the AI regulatory framework. The second section introduces types of bias cognitive, gendered, and algorithmic and explains how socio-technical processes translate human prejudices into modelled outcomes. The third section traces how algorithmic discrimination emerges across the ML lifecycle (data poisoning, biased labeling, and black-box opacity) and examines concrete impacts across recruitment, promotion, performance evaluation, and workplace monitoring. Four case studies (Amazon, Facebook, Apple Card, and HireVue) connect theoretical mechanisms to empirical instances of disparate outcomes and opacity-related harms. The final section assesses legal and policy strategies for mitigation, evaluating EU-level responses such as risk-based AI regulation, pay-transparency measures, and shifts in burden of proof, alongside recent Italian initiatives addressing AI governance, HR transparency, and gender-equality certification. Drawing on the doctrinal and empirical findings, the thesis proposes operational recommendations for policymakers and HR practitioners (continuous algorithmic auditing, context-sensitive fairness metrics, datasheets/model cards, gender impact assessments, and participatory design) and outlines research priorities including context-aware metrics and intersectional monitoring.| File | Dimensione | Formato | |
|---|---|---|---|
|
Cecchin_Gioia.pdf
Accesso riservato
Dimensione
1.02 MB
Formato
Adobe PDF
|
1.02 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/117515