Who gets represented by Indonesian AI? measuring regional, gender, and sociolinguistic bias in large language models

Mohammad Anis Sumadi (Institut Agama Islam Al-Urwatul Wutsqo)

Indonesian Journal of Computational Language Studies
Vol. 1 No. 2 (2026): Large language models, linguistic diversity, and responsible NLP in Indonesia

Background: Indonesia’s regional, gendered, and sociolinguistic diversity raises a critical question about whether large language models represent Indonesian identities with equal specificity, agency, and legitimacy in computational discourse. Objective: This study aims to examine how Indonesian-facing large language models generate representations of regions, gender markers, occupations, and language varieties under controlled prompt conditions. Method: Using a prompt-based audit design, this study analyses 42 prompt units divided into regional, gender-counterfactual, and sociolinguistic conditions, with coding focused on visibility, specificity, agency, competence, register alignment, semantic stability, and language shifting. Results: The findings indicate that regional representation is uneven: some regions are profiled through professional competence, while others are rendered through generic neutrality, cultural tokenisation, peripheral framing, or national homogenisation. Gendered outputs show partial professional parity, but male-coded subjects receive stronger leadership and technical authority, whereas female-coded subjects are more often associated with care, affect, and relational labour. Implication: Sociolinguistic robustness is strongest in formal Indonesian, more adaptive in colloquial Indonesian, and less stable in local-language conditions, where semantic drift, code-mixing, and defaulting to Indonesian appear. Novelty: This study contributes an intersectional audit framework that reframes Indonesian AI bias as a problem of regional visibility, gendered agency, and sociolinguistic legitimacy across culturally stratified AI systems in multilingual Indonesia today.