Critical ReviewInterdisciplinary

Algorithmic Discrimination Is Not a Bug: How Online Targeting Reproduces Racial Inequality

A grounded-theory study reframes algorithmic discrimination not as a technical glitch to be patched but as the digital reproduction of historical racial structures โ€” with online targeted advertising as the mechanism.

By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.

Algorithmic Discrimination Is Not a Bug: How Online Targeting Reproduces Racial Inequality

The standard narrative about algorithmic bias runs something like this: engineers build systems, biased data creeps in, unfair outcomes emerge, and the fix is better data or smarter fairness constraints. It is a story of technical error โ€” regrettable, but ultimately solvable within the engineering paradigm that created the problem.

Bui, McIlwain, Olojoce, and Chang reject this narrative entirely. In a 2025 study published in Information, Communication & Society, they argue that algorithmic discrimination is not a malfunction to be debugged but a digital reproduction of socio-historical racial structures (Bui et al., 2025). The distinction matters enormously. If bias is a bug, you patch the code. If discrimination is structural reproduction, patching the code may leave the architecture of inequality untouched.


The Research Landscape: From Fairness Metrics to Structural Critique

The past decade has produced an enormous body of work on algorithmic fairness. Computer scientists have developed increasingly sophisticated definitions of fairness โ€” demographic parity, equalized odds, calibration โ€” and proposed technical interventions to satisfy them. Simultaneously, social scientists and legal scholars have questioned whether mathematical definitions of fairness can capture the lived reality of discrimination.

This paper positions itself firmly in the structural-critique tradition. Rather than asking how to make algorithms fairer by some technical metric, the authors ask a prior question: what is algorithmic discrimination, ontologically? Their answer draws on the sociology of race and the history of discriminatory practices to reframe the entire conversation.

The specific domain they examine is online targeted advertising โ€” the vast machinery through which platforms sort users into segments and deliver differentiated content, offers, and opportunities based on inferred characteristics. Investigative reporting and academic audits have repeatedly documented racial disparities in ad delivery for housing, employment, and credit โ€” categories where discrimination is illegal under U.S. civil rights law.


What the Paper Does

Bui et al. employ grounded theory methodology to develop their conceptual framework. Rather than testing a pre-specified hypothesis, grounded theory builds theory inductively from qualitative data, through iterative coding and constant comparison. The authors analyze online targeted advertising as what they term a discrimination device (Bui et al., 2025).

The core conceptual move is to connect two processes that are usually discussed separately. First, past racial trend-based pattern recognition: algorithms learn from historical data that encodes decades (or centuries) of racial stratification. When a system trained on such data identifies "patterns," it may be detecting not individual preferences but the sediment of structural racism. Second, selective resource exposure: targeted advertising determines who sees which opportunities โ€” job listings, loan offers, educational programs, housing advertisements. The combination, the authors argue, creates a mechanism through which historical inequality is not merely reflected but actively reinforced in digital environments (Bui et al., 2025).


Critical Analysis

<
#ClaimSourceVerdict
1Algorithmic discrimination should be conceptualized not as technical error but as digital reproduction of socio-historical racial structuresBui et al. (2025), abstractA strong theoretical reframing; the persuasiveness of this claim depends on the granularity of the grounded-theory analysis in the full text, which the abstract does not detail
2Online targeted advertising functions as a discrimination deviceBui et al. (2025), abstractThe characterization aligns with documented cases of discriminatory ad delivery, though the abstract does not specify which platforms or advertising systems were analyzed
3Pattern recognition based on past racial trends combined with selective resource exposure reinforces inequalityBui et al. (2025), abstractThe mechanism is plausible and consistent with prior scholarship on feedback loops in algorithmic systems; however, the abstract does not describe the empirical data from which this framework was inductively derived

The paper's central contribution appears to be conceptual rather than empirical in the narrow sense. By framing targeted advertising as a "discrimination device" โ€” a term that evokes the sociological concept of dispositifs โ€” the authors attempt to shift the locus of analysis from individual algorithmic decisions to the systemic role that advertising infrastructure plays in distributing opportunity along racial lines.

This reframing carries significant implications: standard remedies like debiasing training data or adding fairness constraints may be necessary but insufficient if the underlying architecture perpetuates inequality by design.

At the same time, the structural-reproduction framework raises methodological questions. Grounded theory's strength is its capacity to generate rich, contextually grounded concepts. Its limitation, at least in some formulations, is the difficulty of specifying the boundaries of the claims it produces. The abstract does not indicate the specific data sources โ€” interviews, platform documentation, advertising datasets, case studies โ€” from which the theory was constructed, making it difficult to assess the empirical grounding from the abstract alone.


Open Questions

  • Mechanism specificity: The paper identifies pattern recognition and selective exposure as the twin engines of digital reproduction. But how precisely do these mechanisms operate across different advertising platforms with different architectures? The degree of specificity matters for both scholarly critique and regulatory intervention.
  • Agency and resistance: If targeted advertising reproduces racial structures, what forms of resistance or disruption are possible? The structural-reproduction framework may risk presenting inequality as so deeply embedded as to be intractable โ€” a theoretical cul-de-sac that undermines the political urgency the authors presumably intend.
  • Intersectionality: Racial discrimination in advertising intersects with gender, class, geography, and other axes of inequality. Does the framework account for these intersections, or does it treat race as an isolated variable?
  • Comparative context: The socio-historical structures the authors describe are presumably specific to particular national contexts (most likely the United States, given the author affiliations). How does the framework travel to contexts with different racial histories โ€” Brazil, South Africa, India, or European nations with distinct patterns of ethnic stratification?
  • Regulatory implications: If the "bug" framing implies technical fixes and the "structural reproduction" framing implies systemic transformation, what does the latter actually demand in terms of policy? Advertising bans? Platform redesign? Reparative data practices? The distance between structural diagnosis and actionable prescription remains a persistent challenge in critical scholarship.

  • Closing Reflection

    Bui et al. ask us to stop treating algorithmic discrimination as an engineering problem with an engineering solution. The request is uncomfortable precisely because the engineering framing is convenient โ€” it localizes responsibility, suggests tractable interventions, and preserves the legitimacy of the systems that produce discriminatory outcomes.

    The structural-reproduction lens offers no such comfort. It suggests that the algorithms are working exactly as the social order they were trained on would predict. The patterns they detect are real patterns โ€” patterns of exclusion, segregation, and unequal resource distribution that predate the internet by centuries. The question is not whether the algorithm is accurate but whether accuracy in the service of a discriminatory status quo constitutes a form of discrimination in itself.

    That question has no technical answer. It belongs to the domain of politics, ethics, and collective choice โ€” which may be precisely the point.


    References (1)

    Bui, M., McIlwain, C., Olojoce, S., & Chang, H.-C. H. (2025). Algorithmic Discrimination as Historical Reproduction. Information, Communication & Society.

    Explore this topic deeper

    Search 290M+ papers, detect research gaps, and find what hasn't been studied yet.

    Click to remove unwanted keywords

    Search 6 keywords โ†’