Recent work in computational linguistics has begun testing versions of the linguistic relativity hypothesis within large language models (LLMs), examining how differences in linguistic structure correlate with differences in model reasoning and representation. In this paper, I argue that systematic biases in LLMs, including culturally contingent associations and differential prediction patterns across languages and dialects, constitute empirical support for weak Whorfian hypotheses.
In this paper, we argue that linguistic bias in large language models constitutes a distinctive form of proxy discrimination in which language encodes, rather than merely reflects, social categories. Drawing on philosophy of language, sociolinguistics, and AI fairness literatures, we show how linguistic variation complicates standard causal and correlational accounts of proxy discrimination.
This paper engages philosophical traditions that treat language as a framework for worldmaking to examine how linguistic loss and colonial imposition reshape the contours of being and understanding. I argue that linguistic domination operates not only as cultural erasure but as a reconfiguration of what is perceived, valued, and known.
in Contemporary Issues in Philosophy, Politics, and Economics, forthcoming
This paper examines how processes of meaning-making intersect with algorithmic governance, arguing that linguistic interpretation is central to how computational systems produce and enforce social order. My central claim is that algorithmic systems function through linguistic abstraction: they strip socially situated expressions, categories, and judgments of their contextual meanings and recode them as quantifiable indicators—of compliance, risk, value, etc., as in "creditworthy," "high risk," "toxic," etc.—a process that has yet to receive sustained epistemic or political analysis.
This paper advances a metaphilosophical argument about method in Black philosophy. We distinguish the question of what makes philosophy Black from the question of what counts as philosophy, and argue that conflating these questions produces a misleading methodological monism. We defend a disciplined pluralism according to which methods are justified by problem-fit rather than by allegiance to a single tradition.
This project investigates how the geometric structure of LLM embedding spaces encodes social meaning and linguistic variation—what sociolinguists term indexical fields. The project bridges computational linguistics, philosophy of language, and sociolinguistics to examine how models represent racialized, gendered, and classed associations with linguistic forms. Building on empirical findings that LLMs penalize non-standard dialects (e.g., African American English) and pragmatic styles such as hedging, our research reinterprets these disparities through the lens of indexicality: we hypothesize the LLM embedding spaces represent quantitative representations of indexical fields. Our aim is to develop a scalable computational method for mapping social meaning in LLMs, making indexical field theory operationalizable for AI fairness research.
I'm still working on getting my dissertation portioned out into papers. In the meantime, here's an abstract:
Linguistic profiling—the practice of inferring social identities and evaluative traits from linguistic cues—constitutes a pervasive mechanism of social stratification and discrimination. Under the idealizing assumptions of traditional philosophy of language, however, linguistic meaning is treated as largely independent of listener interpretation. On this view, linguistic profiling would fall outside the scope of meaning per se, since it arises from listener-side inferences rather than from semantic content or speakers' communicative aims. This dissertation challenges that assumption by situating linguistic profiling within a broader class of information-exchange phenomena, enabling a single modeling architecture to capture the diverse manifestations of social meaning. Specifically, I argue that linguistic profiling is only one facet of a broader phenomenon of social identity signaling through language, a continuum that stretches from intentional practices like code-switching and style-shifting on one end, to linguistic profiling itself on the other. To account for this spectrum, I suggest a unified theoretical and formal account integrating sociolinguistics, philosophy of language, and probabilistic modeling. I introduce the Linguistic Profiling Game, a Rational Speech Act–inspired framework for modeling both intentional and unintentional identity signaling as Bayesian inference over utterance, identity, and trait spaces.