<![CDATA[This study evaluates socioeconomic bias in three large language models (LLMs) that support Indonesian Nusantara, IndoGPT, and SEA-LION using an ethical probing approach. A total of 100 short narrative prompts (4–11 words) were compiled to represent issues of poverty, informal employment, access to education, and regional contexts. Each model output was analyzed using five key indicators: emotional valence, stereotypes, narrative themes, framing, and deontic indicators. The results show that all three models tend to produce neutral responses, especially SEA-LION, which has the highest proportion of neutral responses. However, stereotypes still appeared at almost the same level across all models, indicating that a neutral tone does not guarantee bias-free output. IndoGPT showed the highest use of normative language, while Nusantara more often displayed structural framing and empathetic nuances. In contrast, SEA-LION was the most stable in maintaining neutrality without eliminating implicit stereotypical tendencies. These findings confirm that socioeconomic bias in Indonesian-language LLMs still occurs subtly through deterministic narratives, generalizations, and framing that normalizes the vulnerability of low-income groups. This study provides an initial overview of the direction of generative bias in Indonesian LLMs and highlights the need for broader dataset development, stricter annotation methods, and continuous evaluation for the development of fairer models.]]>
