Fairness Across Fields: Comparing Software Engineering and Human Sciences Perspectives

1. Problem Statement

As artificial intelligence (AI) and machine learning (ML) systems increasingly influence critical social domains (healthcare, hiring, finance, public safety), the concept of fairness has become a central quality attribute in software engineering. However, a significant disciplinary gap exists:

• Software Engineering (SE) predominantly treats fairness as a technical, formal, and statistical property. It focuses on measurable metrics, bias mitigation, and algorithmic verification, often abstracting away the social, historical, and power dynamics in which these systems operate.
• Human Sciences (HS) (sociology, law, philosophy, etc.) conceptualize fairness as a context-dependent, sociopolitical construct. It is viewed through the lens of structural inequality, power relations, historical injustice, and collective deliberation, arguing that fairness cannot be reduced to a single metric.

The core problem is that SE approaches often fail to account for the broader socio-technical implications of AI, potentially reinforcing existing inequalities under the guise of "neutral" technical solutions. This study seeks to bridge this gap by systematically comparing how fairness is defined, operationalized, and challenged in both fields.

2. Methodology

The authors conducted a comparative analysis using two complementary review methodologies to synthesize literature from both domains:

• Software Engineering Review: A Systematic Tertiary Review of 45 existing systematic reviews and literature surveys published between 2017–2025. This approach mapped the state-of-the-art in SE regarding fairness definitions, metrics, and mitigation techniques.
• Human Sciences Review: A Rapid Review of 25 articles from the humanities and social sciences (published 2016–2025). This method was chosen to capture fast-evolving theoretical and empirical discussions on algorithmic fairness, focusing on conceptual frameworks rather than technical metrics.

Data Analysis:

• Search Strategy: Specific search strings were used across databases (Google Scholar, ACM, IEEE, Scopus, etc.) targeting "software fairness," "algorithmic bias," and "human sciences" perspectives.
• Extraction & Synthesis: Data was extracted based on five Research Questions (RQs):
  1. How is fairness defined/conceptualized?
  2. What metrics/criteria are used? (Note: RQ2 was deemed inapplicable for HS).
  3. What social benefits are associated with fair practices?
  4. What approaches/methods are proposed to promote fairness?
  5. What limitations and challenges are identified?
• Validation: Thematic synthesis was performed using qualitative coding, with inter-rater reliability ensured through biweekly discussion sessions between reviewers.

3. Key Contributions

• Cross-Disciplinary Mapping: The first study to systematically juxtapose the operational definitions of fairness in SE against the critical, normative definitions in the human sciences.
• Identification of Conceptual Divergence: The paper explicitly delineates how SE reduces fairness to distributional outcomes and statistical parity, whereas HS views it as a structural and relational process rooted in power and justice.
• Critique of Technical Reductionism: The authors argue that treating fairness solely as a computable metric obscures the material and historical conditions that generate discrimination, leading to "fairness debt."
• Framework for Integration: The study proposes a "Socio-technical" perspective that integrates measurable rigor with contextual ethical awareness.

4. Key Results

A. Conceptualization of Fairness

• Software Engineering: Defines fairness as the absence of discrimination in result distribution. It is treated as a formal property (mathematical/statistical) that must be verifiable. Common definitions include Group Fairness, Individual Fairness, and Causal Fairness.
• Human Sciences: Defines fairness as a sociopolitical and historically situated construct. It emphasizes that unfairness is diffused through institutional infrastructures and power relations. Fairness is not just about the output but about the process of power, justice, and participation.

B. Metrics and Evaluation

• Software Engineering: Relies heavily on quantitative metrics derived from binary classification (e.g., Demographic Parity, Equalized Odds, Predictive Parity, Counterfactual Fairness). The focus is on detecting and mitigating statistical disparities between groups.
• Human Sciences: Rejects the sufficiency of metrics. Argues that a system can be statistically "fair" while still being ethically unjust if it reproduces structural hierarchies.

C. Proposed Approaches

• Software Engineering: Focuses on technical interventions:
  • Pre-processing (data cleaning/reweighting).
  • In-processing (algorithmic constraints during training).
  • Post-processing (adjusting model outputs).
  • Use of fairness toolkits and "Human-in-the-loop" systems.
• Human Sciences: Focuses on structural and participatory interventions:
  • Embedding multiple perspectives (feminist, decolonial) in design.
  • Investigating intersecting structures of oppression.
  • Participatory governance and collective agency.
  • Ethical Refusal: The argument that some technologies (e.g., surveillance) should not be built at all, regardless of fairness adjustments.

D. Limitations and Challenges

• Software Engineering: Struggles with the trade-off between fairness and performance, the lack of consensus on which metric to use, the difficulty of handling intersectionality, and the "black box" nature of models.
• Human Sciences: Critiques SE for reducing complex social problems to technical fixes, failing to address root causes of inequality, and reproducing existing power structures through "neutral" algorithms.

5. Significance and Implications

The study concludes that while software engineering provides necessary tools for operationalizing fairness, it is insufficient on its own. The "technical" view often renders invisible the social realities (geography, violence, historical context) that shape algorithmic impact (e.g., the Uber driver case in Rio de Janeiro).

Implications for Practice:

• Beyond Metrics: Developers must move beyond checking statistical parity to understanding the sociotechnical context of deployment.
• Process over Product: Fairness should be viewed as an ongoing, negotiated process involving diverse stakeholders, rather than a static system property to be verified.
• Interdisciplinary Collaboration: Effective AI governance requires integrating SE's technical rigor with HS's critical reflection on power and justice.
• Ethical Refusal: Practitioners should consider whether certain technologies should be developed at all, rather than assuming all systems can be made "fair."

Future Work:
The authors plan to extend this analysis by interviewing software practitioners to create a tripartite comparison between SE literature, SE practice, and Social Science theory, aiming to develop a more holistic framework for responsible AI.