Full e-thesis link is provided below. If the link is broken, the thesis can also be found in the University of Nottingham e-Theses Repository.
Abstract:
This thesis explores how fairness and transparency are perceived by key stakeholders in algorithmic credit scoring systems. As financial institutions increasingly adopt algorithmic tools to make lending decisions, concerns have emerged about whether these systems are fair and transparent. This research investigates how fairness is understood and experienced by borrowers and model developers, and how explanation styles and user factors influence fairness perceptions.
The study follows a mixed methods approach. The first aspect involved focus groups with borrowers to understand their views on fairness and transparency. Borrowers highlighted concerns about rigid eligibility criteria, limited understanding of credit scoring processes, and the absence of human engagement. Fairness was associated with empathy, consideration of individual context, and opportunities to improve future outcomes. Transparency was considered critical, especially when paired with explanations that were clear and actionable. Counterfactual explanations-those offering steps to improve creditworthiness-were viewed most positively.
The second qualitative aspect involved interviews with model developers. Developers emphasized procedural fairness, focusing on objectivity, consistency, and minimizing bias through automation. While they recognized the importance of transparency and explanation, they also noted the challenges of balancing clarity with technical accuracy and protecting proprietary information. They expressed hesitation about including human oversight or subjective data, citing concerns about reintroducing bias and reducing efficiency.
The quantitative aspect examined how explanation formats (text, visual, combined), AI literacy, and trust in AI influence fairness perceptions. Participants who received both text and visual explanations rated the system as fairer across procedural, distributive, and interpersonal dimensions. AI literacy was positively linked to procedural fairness perceptions. Trust in AI was also a significant independent predictor of fairness across all three dimensions. However, it did not moderate the relationships between either AI literacy or transparency and any fairness dimension, contrary to expectations.
The findings reveal both convergence and divergence between borrowers and developers. While both value transparency and explanations, borrowers emphasize distributive and interpersonal fairness, whereas developers prioritize procedural consistency and model performance. These differences highlight the importance of designing credit scoring systems that are not only technically robust but also aligned with users’ values and expectations.
This thesis makes several contributions. It introduces a comparative stakeholder perspective into fairness research, offering new insights into how fairness is defined and prioritised in practice. It extends the fairness discourse by incorporating interpersonal fairness, a less examined dimension in algorithmic systems. The quantitative study is among the few to jointly examine explanation formats, AI literacy, and trust in AI, revealing how these factors interact to shape fairness perceptions. These findings have practical implications for system design, theoretical implications for fairness modelling, and policy relevance for the development of fairer, more transparent and user-aligned decision-making approaches in financial services.
School of Computer Science
Supervisors: Steve Benford, Bernd Stahl, Joy Egede
Keywords: fairness, transparency, algorithmic credit scoring systems, credit scores
