Understanding Algorithmic Fairness: An Investigation of How Stakeholders Construct Algorithmic Fairness in the HR Context

Abstract

Algorithmic fairness has gained increasing scholarly and public attention, as numerous examples show that algorithmic decision making (ADM) systems can introduce new forms of unfair treatment. This thesis responds to calls for a sociotechnical approach by analysing how fairness is constructed by stakeholders involved in developing, implementing and using ADM systems in Human Resources (HR). The research comprises four interconnected studies.

The first manuscript provides an organising review of the algorithmic fairness literature, examining which stakeholders are considered and how fairness is conceptualised from different perspectives. It shows that most research focuses on decision subjects, with far less attention to developers and organisational users, revealing that notions of fairness vary across stakeholder groups.

The second manuscript presents a thematic analysis of software supplier websites, particularly people analytics vendors, to explore how fairness is constructed publicly. Fairness is acknowledged but often undefined or decontextualised, indicating limited transparency.

The third manuscript draws on interviews with HR professionals, software developers, people analysts, and AI and HR consultants to examine stakeholder understandings of fairness and the factors shaping them. Findings show that fairness is constructed through social and technical dimensions, with both shared and conflicting interpretations.

The fourth manuscript addresses a key constraint: HR professionals were not yet using ADM systems. Interviews investigate their concerns and the factors informing them, identifying barriers to implementation and showing how fairness, alongside broader concerns, shape decision making and willingness to adopt ADM technologies.

Overall, the thesis argues that algorithmic fairness cannot be addressed through a one-size-fits-all approach, highlighting how stakeholder perspectives shape its construction and influence adoption.