On Human Factors in Machine Fairness: Essays in Behavioral Economics

Abstract

This dissertation examines the human factors shaping the fairness of algorithmic systems and the distributional consequences of digital technologies. While the design and governance of artificial intelligence (AI) systems are often guided by normative principles, their real-world impact ultimately depends on how individuals and society perceive, interpret, and adopt them. Combining behavioral economics research with the current AI ethics literature, the thesis shows that human behavior can undermine well-intentioned fairness interventions and amplify inequalities. It draws on three complementary empirical approaches: (i) a large-scale deliberation experiment with UK participants on public approval of AI in the public sector, combined with natural language processing (NLP) of transcripts to study attitude formation; (ii) an online lab-style hiring experiment that identifies the causal effect of fairness interventions in algorithmic recommendation tools on their adoption by human decision-makers; and (iii) analysis of representative German household panel data to measure socioeconomic disparities in both actual digital skills and confidence in these skills. The results show that public approval of AI is fragile: it can be quickly raised through favorable information but is just as quickly eroded through more in-depth deliberation, whereas public opposition remains stable. Fairness interventions in algorithmic recommendation tools can backfire by reducing algorithm adoption and thereby reintroducing discrimination at the human decision-maker level. Finally, digital skills and confidence are unequally distributed, potentially reinforcing existing labor-market disparities.These findings highlight the need to incorporate behavioral mechanisms into the design and governance of emerging technologies to ensure that intended objectives are achieved in practice.