PaperlayerView exact revision on arXiv

Source-linked AI summary

A Sociotechnical View of Algorithmic Fairness

Mateusz Dolata, Stefan Feuerriegel, Gerhard Schwabe

arXiv:2110.09253v1cs.CYcs.LGstat.ML

TL;DR

Automated decision-making can have harmful biases, while algorithmic fairness rests on assumptions that remain insufficiently unpacked. The paper examines fairness through a sociotechnical lens and concludes that technical approaches cannot guarantee fair outcomes at scale because systems are dynamic and complex.

  • Problem

    Biases in automated decision-making have negative implications, but assumptions underlying algorithmic fairness remain unpacked as a black box.

  • Method

    The paper uses problematization to facilitate more influential management and organizational research and considers how social and technical components change over time.

  • Results

    Technical approaches cannot guarantee a fair outcome at scale because the system is dynamic and complex, and human values have not been articulated so algorithms can simply obey them.

  • Takeaways & Limitations

    Information systems research should pursue further research into sociotechnical algorithmic fairness with sensitivity to societal progress.

  • Takeaways & Limitations

    It remains hard to evaluate the fairness of technological fairness constraints in practice, especially in a complex and fuzzy context.

Abstract

from arXiv · show

Algorithmic fairness has been framed as a newly emerging technology that mitigates systemic discrimination in automated decision-making, providing opportunities to improve fairness in information systems (IS). However, based on a state-of-the-art literature review, we argue that fairness is an inherently social concept and that technologies for algorithmic fairness should therefore be approached through a sociotechnical lens. We advance the discourse on algorithmic fairness as a sociotechnical phenomenon. Our research objective is to embed AF in the sociotechnical view of IS. Specifically, we elaborate on why outcomes of a system that uses algorithmic means to assure fairness depends on mutual influences between technical and social structures. This perspective can generate new insights that integrate knowledge from both technical fields and social studies. Further, it spurs new directions for IS debates. We contribute as follows: First, we problematize fundamental assumptions in the current discourse on algorithmic fairness based on a systematic analysis of 310 articles. Second, we respond to these assumptions by theorizing algorithmic fairness as a sociotechnical construct. Third, we propose directions for IS researchers to enhance their impacts by pursuing a unique understanding of sociotechnical algorithmic fairness. We call for and undertake a holistic approach to AF. A sociotechnical perspective on algorithmic fairness can yield holistic solutions to systemic biases and discrimination.

1. Introduction

The paper argues that algorithmic fairness is not solely technical: unfairness arises from societal, organizational, and technical sources whose structures mutually reinforce one another. It therefore develops a sociotechnical perspective that integrates technical and social efforts to support holistic fairness solutions.

  • Existing view: Technical AF research quantifies bias and seeks to mitigate discrimination, but technical remedies have dominated treatments of unfairness.
  • Problem: Unfairness in algorithmic decision-making has societal, organizational, and technical sources and is reinforced by both social and technical structures.
  • Sociotechnical framing: Algorithmic decision-making intertwines social and technical components because algorithms can support, delude, or provide excuses for human decisions.
  • Implications: A sociotechnical perspective can yield holistic solutions to unfairness and help coordinate technical innovation, political or legal action, and social awareness.
  • Conceptual contribution: The paper defines sociotechnical AF as algorithmic, organizational, and processual means that prevent systematic discrimination across whole decision processes involving algorithms.
  • Research approach: The authors review AF literature, interrogate its technical and social premises, map bias origins in sociotechnical systems, and formulate directions for IS research.

2. Background on Fairness and Algorithmic Fairness

Algorithmic fairness uses mathematical notions to detect and mitigate disparate harm, but fairness remains socially contested and difficult to operationalize. Group- and individual-level approaches each depend on assumptions that constrain their application.

  • Algorithmic fairness seeks to detect, quantify, and mitigate disparate harm across subgroups affected by automated decision-making.
  • Fairness as a Mathematical Construct: Technical approaches reduce fairness to mathematical expressions and integrate them into algorithms as constraints.
  • Fairness as a Mathematical Construct: Group-level fairness relies on predefined sensitive attributes and compares outcomes or prediction errors across protected and unprotected groups.
  • Fairness as a Mathematical Construct: Statistical parity requires equal event likelihoods across groups, whereas equality of accuracy requires equal prediction accuracies.
  • Fairness as a Mathematical Construct: There is no universal operationalization of fairness, all mathematical notions cannot be satisfied simultaneously, and designers receive limited guidance for choosing among them.
  • Fairness as a Mathematical Construct: Algorithmic fairness faces practical constraints because sensitive attributes may be legally unavailable, while individual-level approaches leave relevant attributes underspecified.
  • Fairness as a Mathematical Construct: Individual-level fairness treats similarly situated people similarly, but requires a use-case-specific similarity definition and leaves relevant nonquantitative attributes difficult to formalize.
  • Fairness as a Social Construct: Fairness is also understood socially as continually constructed through justice, distribution, relationships, perceptions, and contextual influences rather than as a fixed definition.

3. Methodology

The study uses problematization and a multidisciplinary literature review to interrogate assumptions in the emerging algorithmic-fairness discourse. It analyzes technical and social approaches to establish a basis for theorizing algorithmic fairness sociotechnically.

  • The authors use problematization to question assumptions in existing algorithmic-fairness studies and develop an informative research agenda for information systems.
  • Literature Analysis: The analysis identified implicit and explicit assumptions using Alvesson and Sandberg’s categories to support theorizing algorithmic fairness sociotechnically.
  • Literature Collection: The literature review combines algorithmic-fairness conference articles with a query-based search of multidisciplinary outlets.
  • Literature Analysis: Articles were classified by technical or social approach, component focus, scope, methodological paradigm, and research questions.

4. Problematizing Algorithmic Fairness

The literature review identifies divergent technical and social assumptions in algorithmic fairness research, finding no shared coherent agenda. It argues that treating AF as a sociotechnical construct can reconcile these approaches.

  • Review approach: A systematic literature review classified common algorithmic-fairness assumptions as either technical or social orientations.The analysis identified articles exemplifying these assumptions and used them to problematize AF research.
  • Literature-level problem: The reviewed articles lacked a shared, coherent agenda, although their assumptions could even contradict one another.The authors state that not all assumptions coexist in every paper and do not portray the approaches as irreconcilable camps.
  • Sociotechnical direction: The authors argue that AF should be understood sociotechnically because system characteristics emerge from interactions among technical parts, social parts, and context.This framing treats human-algorithm ensembles as mutually influencing components rather than independent elements.
  • Technical assumptions: Technical AF often assumes that biases can be identified in advance and resolved through mathematical formalization without generating new biases.This engineering orientation treats fairness as a problem with a human-made technical solution.
  • Technical assumptions: Technical AF commonly treats distinct fairness ideals as interchangeable and selects among them without consistent attention to selection procedures or long-term social consequences.Participatory and survey-based approaches move selection beyond designers, but their implications and sampling-related biases remain insufficiently considered.
  • Technical assumptions: Translating abstract fairness ideals into strict mathematical metrics can obscure value assessments, context, ethical ambiguity, and unexpected consequences.The authors call this the translation assumption.
  • Technical assumptions: Quantifying fairness can narrow attention to distributive justice even though system usability, dignity, and procedural or interactional justice also matter.The facial-recognition example illustrates how technical failures can affect people beyond distributional disparities.

5. A Sociotechnical Perspective on Algorithmic Fairness

The paper positions algorithmic fairness as a sociotechnical phenomenon because fairness outcomes depend on reciprocal interactions among humans, technologies, information, and broader contexts. This perspective rejects purely technical or social accounts and evaluates fairness alongside other system goals.

  • A sociotechnical account rejects the assumption that improving only a technical component will necessarily make the overall information system fairer.
  • Fairness is one necessary system goal among multiple potentially interrelated or contradictory objectives, not a sufficient condition for usefulness.
  • Information and data steer interactions between social and technical components but are neither neutral nor independent, requiring critical examination.
  • Algorithmic fairness requires considering humans and algorithms as mutually interacting components of decision-making systems.
  • A successful fairness solution requires the entire sociotechnical system to reach coherence and generate fewer biases, rather than assigning responsibility to one component.
  • Because fairness operationalizations may fit changing environments differently, biases cannot always be specified upfront or assessed only against known inputs.

6. Directions for Sociotechnical Research into Algorithmic Fairness

The paper proposes sociotechnical research directions that locate algorithmic unfairness across the whole system rather than within algorithms or data alone. It calls for holistic evaluation, attention to human–machine adaptation, and research that connects technical artifacts with organizational and societal contexts.

  • The literature review distributes the origins of algorithmic unfairness across the entire sociotechnical system and identifies research directions for addressing discrimination at its source.
  • Research should connect low-quality information to the social and organizational processes through which data are generated and managed.
  • Real-world evaluation is needed because fairness constraints lack sufficient external validation in complex contexts beyond isolated technical studies.
  • Fairness research should examine how humans and machines divide work, form partnerships, and adapt to one another in high-stakes decision-making.
  • Because social and technical components mutually adapt, responsibility for insufficient fairness cannot be reduced to a single component.

7. Implications

The sociotechnical framing has implications for IS evaluation, education, practice, and research. It favors holistic outcome assessment, multidisciplinary work, and sustained attention to changing data, algorithms, and human perceptions.

  • Evaluating sociotechnical outcomes is more appropriate than intervening only in low-level system processes.
  • Algorithmic fairness should be treated as a multidisciplinary endeavor spanning social and technical systems.
  • Singular interventions may work only briefly because data, algorithms, and human perceptions change over time.
  • Fairness-oriented systems can affect organizations, and productivity may increase when social and technical notions of fairness align.
  • Engineering education should combine technological understanding with humanistic, social, and behavioral dimensions, while practitioners should critically assess artifacts.
  • The literature review is limited by its databases, keywords, filtering procedures, and concise article summaries.

8. Conclusion

The paper argues that algorithmic fairness is a dynamic sociotechnical phenomenon whose outcomes depend on reciprocal interactions between social and technical components. It concludes that technical interventions alone cannot guarantee fair outcomes and calls for further sociotechnical research.

  • Algorithmic fairness problems become complex when social and technical components interact reciprocally within a broader environment.
  • State-of-the-art technical approaches cannot guarantee a fair outcome at scale.
  • Social values change through interactions with technology rather than remaining fixed specifications that algorithms can simply obey.
  • The value-alignment debate therefore requires recognizing that values are shaped and negotiated in sociotechnical processes.
  • The paper calls on IS researchers to pursue sociotechnical algorithmic fairness using practical, technical, and socially oriented expertise.

Appendix

The appendix describes a systematic literature review designed to characterize algorithmic-fairness discourse and identify its recurring limitations and assumptions.

  • The review characterized algorithmic-fairness discourse and identified potential limitations and assumptions.
  • The review procedure comprised literature search and selection, classification, and analysis.

Literature Search and Selection

The literature search combined targeted machine-learning conference proceedings with multidisciplinary database and citation searches. After screening and selection, the review assembled conference and multidisciplinary sets of peer-reviewed articles.

  • The multidisciplinary search yielded 149 potentially relevant articles, from which 69 were selected for further processing.
  • 9,392 articles from four machine-learning conferences published between January 2017 and December 2020 were screened for algorithmic-fairness terms.
  • 166 conference articles were retained after title and abstract screening.
  • Backward and forward searches from the selected multidisciplinary articles added 45 articles, producing a multidisciplinary set of 114 articles.
  • The final review basis comprised 280 peer-reviewed articles: 166 in the conference set and 114 in the multidisciplinary set.

Literature Classification

The classification analyzed the literature across fairness perspective, IS component, methodological paradigm, and scope. The review found that technical perspectives and engineering approaches dominated, while no article fully adopted the paper’s sociotechnical perspective.

  • The literature was classified by fairness perspective, IS component, methodological paradigm, and scope.
  • No reviewed article adopted the sociotechnical perspective defined by mutual interdependency, joint optimization, and equivalency between technical and social components.
  • 210 articles took a technical perspective, compared with 70 taking a social perspective.
  • More than 50 percent of studies focused on engineering new approaches to address algorithmic bias.
  • 203 articles made generic claims, while 77 were domain-specific, including work in health, criminal justice, and loan allocation.

Literature Analysis

The literature analysis classified articles into clusters and mapped their assumptions to dominant technical or social perspectives on fairness. The paper then used these patterns to distinguish technical and social assumptions.

  • Articles were examined in clusters ranging from large technology-focused groups to medium broader-context groups and individual cases.
  • The classification compared technical perspectives on fairness with social perspectives involving broader contexts or social subsystems.
  • The authors identified typical assumptions for the clusters and concluded that these assumptions corresponded to the fairness perspective dominating each paper.
  • The paper’s structure differentiates assumptions associated with technical and social perspectives on fairness.
  • The analysis grouped similar or overlapping assumptions to present a comprehensive account of the literature.

Analyzed Articles

The analyzed-article set includes studies spanning technical and social perspectives on algorithmic fairness, with many entries addressing engineering, technology, and broader social concerns.

  • The article inventory is organized under a fairness-perspective classification.
  • Technical and engineering-oriented entries include fair classification, regression, hierarchical clustering, and clustering without over-representation.
  • Other entries address social or broader concerns, including power, racial categories, political philosophy, and algorithm-assisted decision making.
  • The inventory also includes work on fairness in recommendation, allocation, ranking, graph embeddings, generative modeling, and decision making.
Loading 2110.09253v1…