UC Santa Barbara

Algorithms are capable of bias; certain algorithms have been shown to disadvantage certain users or data over others in a way that is either morally problematic or unintended as a function of that algorithm. Not only are some algorithms actually biased, but people also sometimes perceive them to be biased (more or less accurately). Existing approaches to studying algorithmic bias typically follow a two-step process: first identifying bias within specific systems, then examining its effects on users. However, both steps face significant limitations.First, algorithmic systems differ in substantive ways, and are constantly updated, making findings difficult to replicate across time and platforms. Moreover, this approach treats each algorithmic system as unique, limiting our ability to develop generalizable knowledge about how bias operates across different technological contexts. The result is an accumulation of findings about specific systems rather than theoretical understanding of underlying dynamics.Additionally, this two-step approach often ignores the effects that perceptions of algorithmic bias may have on user responses. For instance, when users believe that social networking site (SNS) feeds are biased they may change how they engage with SNS algorithms, regardless of whether the underlying algorithm itself is biased. Conversely, not perceiving bias when the feeds are biased can lead to misinformed and unjustified opinions, decisions, and behavior.Evidence suggests that algorithmic bias often parallels patterns long documented in traditional media. For example, Noble's (2018) research on search algorithms that return stereotypical images for searches involving women of color mirrors decades of scholarship on stereotypical portrayals in traditional media (e.g., Mastro & Stamps, 2018). Similarly, algorithmic systems that associate Black-sounding names with criminal advertisements (Sweeney, 2013) reflect the same representational biases found in news coverage (Dixon & Linz, 2000). These parallels suggest not only that algorithmic bias manifests in familiar patterns, but more importantly, that the theoretical frameworks developed to explain traditional media effects remain conceptually valid in algorithmic contexts. The underlying psychological and social processes that drive media selection and effects do not fundamentally change simply because the delivery mechanism shifts from human editors to algorithmic curation.Despite these parallels, there are differences that distinguish algorithmic from traditional media contexts. Algorithms do not just present real (or perceived) biased content in a static way; they shape what content users encounter based on their prior behaviors and preferences. This creates a feedback loop where user choices, algorithmic curation, and bias perceptions interact dynamically. Understanding this dynamic requires theoretical frameworks that account for both user agency in content selection and the algorithmic constraints on that selection. One such theory is selective exposure.Originally developed to explain how people choose attitude-consistent information environments, selective exposure takes on new dimensions in algorithmic contexts. Three key considerations make this theory particularly relevant. First, algorithms alter media choice: where traditional selective exposure involved users actively seeking out preferred sources, algorithmic systems now automate much of this selection process, meaning selective exposure operates both through explicit user choices and through algorithmic inferences about user preferences. Second, user perceptions of algorithmic bias likely influence selective exposure behaviors in ways that differ from traditional media; when users believe an algorithm is biased against their viewpoint, they may attempt to train or curate the algorithm through strategic engagement rather than simply switching to an alternative source. Third, algorithmic awareness may moderate selective exposure effects, as users who understand that algorithms learn from their behavior may be more deliberate in their engagement patterns.This dissertation addresses these dynamics across four chapters. Chapter I develops the conceptual vocabulary needed to study user responses to algorithmic bias, defining algorithmic bias and its perceived counterpart, reviewing related constructs, and explicating Strategic Algorithm Engagement: the deliberate, goal-directed behavioral capacity through which users attempt to shape what algorithmic systems recommend, prioritize, or display to them. Chapter II translates this conceptual work into an empirical measurement instrument, reporting the development and validation of the Strategic Algorithm Engagement Scale across three studies. The resulting four-item scale demonstrates good model fit, strong factor loadings, and full measurement invariance across sex, education, income, and political identification, and is empirically distinct from algorithmic awareness. Chapter III tests these ideas empirically in a case study of partisan SNS feed curation, using behavioral data donated from participants' Facebook archives (N = 419) to examine how perceptions of politically motivated algorithmic bias translate into selective exposure behavior, and how Strategic Algorithm Engagement moderates that process. While the study was underpowered for inferential testing due to low rates of political engagement in the archive data, the descriptive results reveal high ideological selectivity among those who did engage with partisan content, and directionally asymmetric associations between perceived bias, strategic engagement, and curation behavior across partisan groups, patterns that point to directions for future work.Taken together, these contributions provide the conceptual vocabulary, measurement infrastructure, theoretical synthesis, and methodological lessons needed to move the study of human responses to algorithmic bias from documenting that bias exists toward understanding how users perceive it and respond to it.