Epistemic Injustice in Generative AI: How Minority Language Systems Reproduce or Resist Digital Exclusion

Generative AI performance disparities can exclude minority-language speakers from digital participation. Drawing on Fricker’s epistemic-injustice framework, we examined five Chinese minority languages using four-way causal mediation decomposition, intersectional analysis, and participatory evaluation. Under the measured mediators and model specification, non-PIE components accounted for an average of 67.5% of observed disparities. Tibetan speech recognition yielded a 43.9% word error rate (WER) for elderly speakers using religious and dialectal language, 6.9 percentage points above the main-effects additive reference. Community evaluation prioritized religious sensitivity, while threshold rules identified culturally salient failures missed by linear weighting. CAML-LLM, integrating cultural knowledge graphs, multimodal alignment, and community-weighted training, reduced WER from 31.2% to 22.7% and improved cultural appropriateness by 22.5%. These findings characterize model-dependent disparities and demonstrate that participatory system design can improve minority-language performance within the evaluated settings.