The Paternalistic Filter: Epistemic Injustice and Differential Refusal in LLM-Mediated History Education for Marginalized Romanian Students

Jul 16, 2026

Large Language Models are increasingly used as conversational history tutors, yet their deployment can institutionalize systemic inequalities. This study audits four LLMs via a systematic API evaluation of 1,800 responses about the 1989 Romanian Revolution across five student personas differentiated by ethnicity and socioeconomic tier. Four patterns emerge: differential refusal, epistemic gatekeeping, agency theft, and elite hermeneutics, where safety alignment withholds justification and civic complexity. The results argue for urgent pedagogical auditing to prevent hermeneutical injustice.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Introduction

Related Work

Methodology

Experimental Design and Personas

arXiv :2607 . 11292v1 [ cs .CY] 13 Jul 2026 The Paternalistic Filter: Epistemic Injustice and Differential Refusal in LLM-Mediated History Education for Marginalized Romanian Students Alexis Popovici 1 [0009-0006-7309-6674] ⋆ , Andrei Ionas, cu 1 , and Adrian Marius Dumitran 1 Universitatea din Bucures, ti, Bucharest, Romania [alexis.popovici.123@gmail.com](alexis.popovici.123@gmail.com) , [ionascuandrei320@gmail.com](ionascuandrei320@gmail.com) , [marius.dumitran@unibuc.ro](marius.dumitran@unibuc.ro) Abstract. As Large Language Models (LLMs) are increasingly deployed as conversational tutors, they risk institutionalizing systemic inequalities. This study presents a systematic API audit of four LLMs acting as history tutors, evaluating 1,800 responses regarding the 1989 Romanian Revolution across five student personas varying by ethnicity and socioeconomic tier. We uncover four interconnected patterns of epistemic paternalism: (1) Differential Refusal, where safety-aligned models block 76.7% of educational requests from low-tier students; (2) Epistemic Gatekeeping, evidenced by a 3 × reduction in access to geopolitical complexity (e.g., the contested “coup theory”) for marginalized learners; (3) Agency Theft, a lexical shift where models like LLaMA produce a 5 × higher victimization-to-politics vocabulary ratio for Roma students compared to elite peers; and (4) Elite Hermeneutics, where AI tutors disproportionately withhold epistemic confidence and justification scores from low-resource demographic profiles. We argue that current safety alignment acts as a paternalistic filter, transforming conversational AI into agents of narrative segregation—a manifestation of hermeneutical injustice in Fricker’s [1] sense that demands urgent pedagogical auditing. Keywords: Conversational agents · LLM bias · AI in Education · epistemic injustice · safety alignment · cultural fairness 1 Introduction LLMs are rapidly being adopted as history tutors [2] . Their perceived neutrality makes them attractive for politically complex topics. However, the promise of democratized learning assumes uniform treatment—an assumption this study shows to be empirically false. When an AI tutor scales historical complexity based on a student’s perceived identity, it gatekeeps rather than personalizes. This paper investigates whether ⋆ Corresponding author: [alexis.popovici.123@gmail.com](alexis.popovici.123@gmail.com) LLMs systematically adjust content depth, lexical richness, and narrative framing based on a student’s ethnic and socio-economic profile regarding the 1989 Romanian Revolution. This event is politically contested, relevant to local minorities (the Roma), and sufficiently nuanced to reveal divergent AI framing strategies. Our central claim is that LLM behavior constitutes epistemic paternalism: withholding civic complexity from marginalized learners while replacing it with victimization-centered narratives. This mirrors Fricker’s hermeneutical injustice [1], denying marginalized students the conceptual tools to understand their history as political actors. This study directly addresses the intersections of technology (LLM API auditing), education (history tutoring delivery), and culture (marginalized Romanian ethnic identities), fulfilling the mandatory requirements for culturally-aware pedagogical systems. 2 Related Work Weissburg et al. [2] found LLMs perpetuate stereotypes across 17,000 educational explanations, primarily along socio-economic dimensions. Gupta et al. [3] and Cheng et al. [9] demonstrated that persona-assigned LLMs harbor deep implicit biases, manifesting stereotyped vocabulary when responding as marginalized demographics despite explicit safety tuning. Fricker [1] defines hermeneutical injustice as a gap in interpretive resources that disadvantages groups in understanding their social experience. Recent frameworks [4,5] apply this to generative AI, showing LLMs selectively omit vital agentic con

FAQs

这项研究在评估哪些LLM以及评估了什么教学内容？

研究对四个LLM作为历史辅导者进行系统API审计，评估的教学主题是1989年罗马尼亚革命。实验共分析1800个回复，并在五类学生画像之间对比差异。

研究发现了哪四种相互关联的“认识论家长式”模式？

研究提出四类模式：差异性拒绝、认识论把关、主体性被剥夺（词汇层面的变化）以及精英式诠释（对低资源群体不成比例地减少解释与辩护分数）。这些模式共同表现为对边缘化学习者的认知资源剥夺。

研究如何解释这种现象与“解释学不公”之间的关系？

研究认为，当前的安全对齐机制起到了“家长式过滤器”作用，使对话式AI成为叙事隔离的推动者。其具体表现对应到Fricker所说的解释学不公：边缘化学生被剥夺理解自身历史作为政治行动者所需的概念工具。

Jul 16, 2026

Upload by Evangeline