Work

Towards Fair AI Systems: Programmer Practices and Human Machine Interaction

Year: 2026
Type: dissertation
Abstract: This thesis studies fairness in AI-assisted decision systems as a behavioral and organizational problem that unfolds across the development, deployment, and learning stages of algorithmic systems. Rather than treating bias as a purely technical property of machine learning models, the thesis examines how fairness outcomes are shaped by the choices of algorithm developers, the cognition of human decision makers, and the structure of repeated interaction between humans and machines. Chapter 2 examines the development stage through two behavioral experiments in which participants build and select machine learning models under different organizational conditions. The results show that stringent accuracy targets increase the likelihood of selecting the least fair model, while fairness norm messaging and accountability requirements shift participants toward fairer submissions. Chapter 3 studies the deployment stage through an analytical model in which a rationally inattentive Decision Maker observes a potentially biased Machine signal, forms a Bayesian belief, and decides how much costly additional information to acquire before taking a final action. The analysis shows that awareness of group-specific Machine error rates can improve both fairness and accuracy relative to unawareness, although awareness does not eliminate disparities under all parameter configurations. Chapter 4 extends this framework by introducing an experience-based learning phase in which the DM learns perceived group-specific error rates from repeated feedback before entering the steady decision phase. The results show that asymmetric sensitivity to Machine errors, group representation in the learning environment, and model-card initialization jointly shape the speed and direction of learned beliefs, which then determine whether the human-Machine system mitigates, preserves, or reverses the Machine’s original bias. Across the three essays, the thesis shows that fair AI systems require governance mechanisms that shape developer behavior, disclosure practices that support informed human oversight, and learning environments that calibrate users’ beliefs about algorithmic errors.
Language: English
Open Access status: green
OpenAlex ID: W7204485062
Cites: 155
Cited by:
Related to: 10