Team Diversity Promotes Software Fairness: An Experiment on Fairness-Aware Requirements Prioritization

1. Problem Statement

While fairness in software engineering is increasingly recognized as critical—particularly for AI-driven systems—current research predominantly focuses on algorithmic bias, data quality, and post-deployment auditing. There is a significant gap in understanding how fairness is addressed during the early stages of software development, specifically during requirements engineering and prioritization.

Furthermore, while the theoretical value of team diversity (EDI: Equity, Diversity, and Inclusion) is often asserted, there is a lack of empirical evidence regarding how specific dimensions of diversity (particularly sexual orientation) influence fairness-aware decision-making. Most existing studies focus on gender or race, leaving the impact of LGBTQIA+ representation on technical reasoning largely unexplored. The core problem is whether diverse teams are better equipped to identify, prioritize, and reject requirements that could lead to discriminatory outcomes before they are coded.

2. Methodology

The authors conducted a controlled experiment involving 27 pairs of software engineering students (54 participants total).

• Participants: Students enrolled in project-based courses with prior training in requirements engineering and agile practices.
• Grouping: Pairs were formed randomly. Post-hoc demographic analysis categorized them into two groups:
  • LGBTQ-diverse pairs (13 pairs): At least one participant identified as LGBTQIA+.
  • Non-diverse pairs (14 pairs): Both participants identified as heterosexual.
• Task: Pairs were asked to prioritize 24 user stories for a fictional AI-based financial system (credit risk evaluation). The stories were pre-classified by researchers into three categories:
  1. Fairness-aligned: Features promoting equitable treatment.
  2. Fairness-risking: Features that could introduce bias or exclusion.
  3. Neutral: Features unrelated to fairness.
• Procedure:
  • Pairs discussed each story collaboratively.
  • They assigned a priority score (1 = Do Not Implement, 4 = High Priority).
  • They provided written justifications for their decisions.
  • Demographic data was collected anonymously after the task to prevent bias during pairing.
• Analysis:
  • Quantitative: Descriptive statistics were used to calculate "Pro-Fairness Scores" (prioritizing aligned + deprioritizing risking) and "Misprioritization Scores" (errors in judgment).
  • Qualitative: Thematic analysis was applied to the written justifications to identify reasoning patterns (e.g., ethical considerations, business value, legal compliance).

3. Key Contributions

1. Operationalization of Fairness in Requirements: The study defines and measures "fairness-aware attitudes" specifically within the context of software requirements prioritization, moving beyond theoretical discussions to behavioral metrics.
2. Empirical Evidence on Sexual Orientation Diversity: It provides one of the first controlled investigations linking LGBTQIA+ team composition to fairness-aware reasoning, filling a gap in literature that has largely ignored sexual orientation in software engineering contexts.
3. Behavioral Link between Diversity and Ethics: The study demonstrates that diversity is not just a representational metric but a functional determinant of how teams identify and interpret fairness risks during the design phase.

4. Key Results

The study found distinct behavioral and reasoning differences between the two groups:

• Higher Pro-Fairness Alignment: LGBTQ-diverse pairs achieved a higher composite pro-fairness score (6.92 vs. 6.50 per pair).
• Decisiveness in Rejecting Harm: The most significant difference was in rejecting fairness-risking stories. LGBTQ-diverse pairs deprioritized these stories more consistently (avg. 5.23 per pair) compared to non-diverse pairs (avg. 4.29 per pair).
• Fewer Errors: LGBTQ-diverse pairs made fewer misprioritization errors (2.15 vs. 2.29 per pair), indicating a sharper ability to identify and avoid features that could lead to discriminatory outcomes.
• Reasoning Patterns (Qualitative):
  • LGBTQ-diverse teams: Their reasoning was explicitly driven by Inclusion, Non-discrimination, and Ethical Responsibility. They treated fairness as a guiding principle, often engaging in explicit dialogue to reach consensus on ethical grounds.
  • Non-diverse teams: Their approach was more pragmatic and goal-oriented. They balanced fairness with Project Scope, Business Value, and Legal Compliance. While they considered ethics, it was often reactive (rejecting clearly unethical items) rather than proactive, and they showed more hesitation (neutrality) when rejecting risky stories.

5. Significance and Implications

• Shift in Fairness Assurance: The findings suggest that fairness cannot be solely addressed through technical audits or model adjustments later in the lifecycle. Fairness must be treated as a first-class requirement addressed during the initial elicitation and prioritization phases.
• Diversity as a Risk Mitigation Strategy: Diverse teams, specifically those with LGBTQIA+ representation, act as a "filter" against bias. Their lived experiences allow them to detect potential harms and discriminatory requirements that homogeneous teams might overlook or rationalize away.
• Organizational Practice: Software organizations should view inclusive team composition not just as an HR goal but as a technical quality assurance mechanism. Integrating diverse perspectives early in the requirements phase can prevent the embedding of biased logic into AI systems.
• Future Research: The study calls for further investigation into intersectionality (e.g., combining race, gender, and orientation) and the application of these findings to professional industrial settings, as the current study was limited to student participants.

Conclusion: The paper concludes that team diversity directly enhances the identification and interpretation of fairness issues during requirements analysis. By fostering more reflective and inclusive decision-making, diverse teams contribute directly to the reliability and social responsibility of AI-enabled software systems.