Resource-Ranked Semantic Drift: GenAI Translation, Equity, and Minority Languages in Taiwan

An emerging assumption that generative artificial intelligence may function as a universal translator obscures the unequal treatment of languages within its architecture. This study challenges the myth of AI translation neutrality by investigating the performance of large language models on two severely under-resourced Austronesian languages in Taiwan, namely Amis and Paiwan. Through an analysis comparing ChatGPT-5.5 translations against formally authorized human-translated baselines, the study demonstrates that semantic loss correlates directly with a language's position in the global resource hierarchy. I conceptualize this phenomenon as "resource-ranked semantic drift," a three-layered theoretical model comprising pragmatic thinning, cultural flattening, and epistemic invention. The findings reveal that AI translation of indigenous languages frequently degenerates into hallucinated narratives that overwrite cultural specificity with dominant-language epistemologies. Concurrently, this paper proposes a Translation Equity Framework incorporating cultural specificity, intra-lingual diversity, and community representational sovereignty, contributing to the broader discourse on postcolonial translation ethics and digital language justice.