Poisoning the Discourse: Detecting Implicit Poverty Stigma in U.S. Political Speech Through Masked Language Models

• In Event: Using AI as a Tool for Social Analysis: Prospects and Challenges

Abstract

While existing research has documented explicit anti-poor rhetoric in welfare debates, the implicit associative dimensions of poverty stigma—the linguistic structures that shape perception without overt derogation—remain systematically unmeasured. This paper proposes a computational methodology to fill this gap: drawing on masked language model (MLM) techniques, I develop a two-model comparison design pitting a base BERT model against a domain-adapted "Stigma-BERT" fine-tuned on a systematically constructed corpus of meritocratic-libertarian discourse, with corpus boundaries defined through discourse network analysis and ideological positioning cross-validated through third-party classification services. Comparing token probability distributions over theoretically grounded sentence templates generates ∆-probability scores quantifying how meritocratic-libertarian discourse shifts implicit poverty associations. The fine-tuned model is treated not merely as a classifier but as a computational instantiation of a discourse structure, and implicit poverty stigma is situated as an upstream mechanism in the discourse-to-policy nexus: stigmatizing frames in elite discourse may penetrate legislative speech, shaping welfare debate framing, public opinion, and policy outcomes. Applied to the U.S. Congressional Record as external validation, the design tests whether the meritocratic logic chain—effort determines success; poverty signals insufficient effort; welfare rewards failure—is embedded in American political speech.

Author

• Yunseok Jeong, Korea University