A systematic API audit of four Large Language Models acting as history tutors reveals that safety alignment mechanisms can institutionalize systemic inequalities for marginalized learners. The study evaluated 1,800 responses regarding the 1989 Romanian Revolution across five student personas varying by ethnicity and socio-economic tier.
- Differential Refusal: Safety-aligned models blocked 76.7% of educational requests from low-tier students.
- Epistemic Gatekeeping: Marginalized learners experienced a 3x reduction in access to geopolitical complexity, such as the contested coup theory.
- Agency Theft: Models like LLaMA produced a 5x higher victimization-to-politics vocabulary ratio for Roma students compared to elite peers.
- Elite Hermeneutics: AI tutors disproportionately withheld epistemic confidence and justification scores from low-resource demographic profiles.
The authors argue that these patterns constitute hermeneutical injustice, transforming conversational AI into agents of narrative segregation that demand urgent pedagogical auditing.