Abstract
dc:description.abstractArtificial intelligence is widely promoted as a revolutionary methodological tool, capable of processing information at a scale and speed no human analyst can match. Behind this promise lies an epistemological assumption: that synthetic reasoning, perception, and learning are neutral procedures that improve upon, rather than reproduce, the limits of human cognition. This dissertation argues the opposite. Drawing on the sociology of knowledge — where reasoning is socially constructed (Durkheim and Mauss 1902; Mannheim 2013), perception is grounded in positional stereotypes (Benjamin 2023; Hill Collins 2009), and learning is socialization into exclusive cultural structures (Bourdieu 2020; Collins 2019) — I argue that contemporary AI systems are not neutral methodologies but political epistemologies. The operationalization of cognition in code does not abstract logic, perception, or learning away from social position; it institutionalizes a particular social position as universal and applies it at scale. Through three empirical studies of the most visible AI systems in everyday use — large language models, platform user-inference systems, and recommendation algorithms — this dissertation demonstrates how the technical components of AI design encode politics into the production of knowledge. The dissertation makes three primary contributions. Theoretically, it builds onto the canonical typology of algorithmic bias (Friedman and Nissenbaum 1996) by centering technical bias — the form most resistant to existing debiasing strategies and least examined in sociological literature, which has largely focused on pre-existing and emergent biases. Methodologically, Chapter 2 develops two novel research designs to address the analytic difficulty of studying proprietary black-box systems: hybrid representative sampling, which combines network theory, stratified sampling, and platform activity metrics to draw representative samples of digital ecosystems; and the composite user audit, which adapts the logic of oracle attacks from adversarial machine learning to characterize how recommendation algorithms learn demographically specific models of identity. Empirically, the three studies that follow show technical bias operating across three distinct cognitive functions: reasoning (Chapter 3), perception (Chapter 4), and learning (Chapter 5). Chapter 3 locates the source of large language model bias not in training data alone, where prior scholarship has largely placed it, but in the transformer algorithm itself — the computational architecture at the heart of every contemporary LLM. I argue that transformers are structurally majoritarian: because they learn by identifying and reproducing prevalent patterns in their training data, whatever voice dominates the corpus will dominate the model's outputs. I demonstrate this first through an experiment training transformer models of varying complexity on synthetic datasets with deliberately unequal word frequencies, showing convergence on training-data prevalence regardless of model size. I then scale the argument empirically, using a hybrid representative sample of Reddit's political ecosystem — a documented ChatGPT training source — alongside a Euclidean distance analysis comparing the media trust ratings of 316 intersectional demographic subgroups to ChatGPT's own responses. ChatGPT's political disposition mirrors that of young white male Democrats, the dominant demographic in Reddit's political communities; women and Black users are systematically further from ChatGPT's political sensibility in proportion to their underrepresentation. This is not a problem debiasing techniques like distillation or pruning can resolve, as they treat symptoms while leaving the majoritarian logic of the architecture intact. Chapter 4 asks how platforms actually see users, and finds an answer that unsettles existing theorizations of surveillance capitalism. Where prior scholarship has detailed the social consequences of being surveilled, sorted, and governed by platforms, the technical mechanisms of that surveillance have remained underexamined. Through a meta-analysis of twenty methodology papers on user-characteristic inference and a survey audit of Google's AdTech profiles administered to 1,045 users across forty-eight countries, I find that platforms do not see users as demographic beings with tastes and identities, but as ripples in digital space: IP addresses, device specifications, scrolling metrics, battery decay rates, browser extension installations. These device signals are operationalized into human-legible demographic categories through inferential statistics vulnerable to noise and randomness. The political epistemology of platform vision is not a measurement problem awaiting a more accurate algorithm; it is a problem of interpretation — whose categories, built from whose data, applied to whose lives, and in whose interests. A more precise inference system would not resolve the underlying issue; it would only apply the social defaults more consistently. Chapter 5 demonstrates the political epistemology of recommendation algorithms at the level of cultural production. Where prior literature on algorithmic bias has largely operated within a harm-and-benefit binary, I argue this framing obscures the in-group prejudices that recommendation systems co-produce with users themselves. Using the composite user audit developed in Chapter 2, I recruited 180 participants across six race-by-gender groups to train six separate Instagram Reels feeds through fifteen hours of naturalistic engagement each. Quantitative findings reveal statistically significant differences in self-portrayal perceptions across racial groups: white participants perceived more positive self-portrayals on average, Black participants perceived an even distribution across positive, neutral, and negative representations, and Asian participants perceived a bimodal distribution concentrated at both poles. Qualitative findings specify these differences in kind, revealing that each demographic group encountered a distinct algorithmically co-produced identity caricature legible through existing critical scholarship (Hill Collins 2006, 2009; Hooks 2005; Rottenberg 2018). I develop the concept of algorithmic ambivalence to describe the process by which recommendation systems enforce group-specific identity models that are simultaneously affirming and denigrating, calibrated to the comportment of each demographic group. The political epistemology embedded in these systems requires neither curated data nor intentional design — only users performing identity and an algorithm learning from that performance. Together, these three studies demonstrate that AI is not a discrete methodological tool but a political infrastructure, encoding social stratification into the production of knowledge through the technical mechanisms of algorithm choice, data pre-processing, and learned feature attribution. As artificial intelligence is increasingly integrated into the knowledge infrastructures of education, government, and information access, recognizing it as a political project — not a neutral methodology — becomes paramount. The dissertation closes with a discussion of what this reframing means for ongoing debates in the sociology of AI, critical algorithm studies, and the future of knowledge-producing institutions.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Sociology
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Beauvais, Taylor
- Advisor dc:contributor.advisor
-
- Mijs, Jonathan
Subjects
dc:subject × 6Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53666