Bias Algoritma dan Kegagalan Pragmatik AI dalam Mengidentifikasi Ujaran Kebencian Berbasis Budaya Lokal di Indonesia

Abstract

This study aims to identify the pragmatic failures of Large Language Models (LLMs) and the biases of Anglophone-based AI moderation algorithms in detecting Indonesian hate speech expressed through sarcasm, satire, euphemism, and local cultural metaphors. It also examines the extent to which AI systems understand and interpret the pragmatic meanings within the corpus. This study employs a qualitative descriptive approach with a comparative design. Data were collected through the documentation of hate speech expressions on social media containing elements of local cultural hatred. The data were analyzed using qualitative descriptive methods with pragmatic and thematic approaches. The findings show that all corpus data contain political satire and indirect hate expressed through irony, sarcasm, absurd metaphors, and popular culture wordplay. Testing with Claude AI showed that the system was capable of identifying the data as implicit criticism and recognizing the pragmatic functions of emoticons and contextual meanings in the utterances. However, the analysis also demonstrated limitations in understanding local sociocultural contexts, particularly the metaphors “daun nangka” and “daun sawit,” which were interpreted merely as absurd humor. These findings indicate that AI detection accuracy does not necessarily reflect a deep pragmatic and cultural understanding within the Indonesian context.