Fairness in social machines: a systematic review

Purpose The purpose of the paper is to provide a systematic review of biases in social machines to better understand the general problem of fairness in these systems. It aims to identify and categorize phenomena described as biases toward specific demographic groups, frame them normatively as harmful and relate them to established fairness concepts originally defined for algorithmic systems. Design/methodology/approach The phenomenon of algorithmic bias refers to systematic biases against identifiable demographic groups that occur in automated decisions systems. Such biases have mostly been studied in the context of black-box decision systems built using machine learning (ML). However, similar problems have also been reported in complex socio-technical systems such as Wikipedia and Airbnb, known more generally as social machines, where the observed biases cannot necessarily be attributed to specific automated decision systems. Instead, the biases may emerge as a result of complex processes involving numerous users and a computational infrastructure. To gain a better understanding of fairness in social machines, the authors select a representative sample of social machines from six distinct categories, and systematically review the literature reporting biases in these systems, covering 196 papers. The authors classify the reported bias phenomena, identify the affected demographic groups and relate the phenomena to established notions of harm from algorithmic fairness research. Finally, the authors identify the normative expectations of fairness associated with the different problems and discuss the applicability of existing criteria proposed for ML-driven decision systems. The analysis highlights the conceptual similarity of bias phenomena between algorithmic systems and social machines, allowing for a shared vocabulary to describe and compare phenomena across a broad class of systems. Findings The paper identifies two key biases in social machines: representational harm, from underrepresentation or biased portrayal of disadvantaged groups, and allocative harm, from unfair decision processes, measurable via metrics like demographic parity. Gender bias is prevalent and easier to detect due to explicit markers, offering insights for identifying other biases. Unique biases arise from user categorizations, creating unintended discrimination linked to protected characteristics. These biases result from complex user interactions, not isolated algorithms. Addressing them requires redesigning social machines, focusing on computational infrastructure and interaction norms, such as visibility settings, to mitigate harmful outcomes. Originality/value The paper’s originality lies in its systematic review of biases in social machines, offering a novel perspective on fairness in these systems. Unlike prior studies focusing solely on algorithmic fairness, this work examines the broader socio-technical interactions within social machines, identifying biases that emerge from user interactions and design choices. By linking these biases to established fairness concepts like demographic parity and representational harm, the paper bridges the gap between algorithmic fairness and social dynamics.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.