<![CDATA[AI assistants often give wrong answers with high confidence, especially when answering in regional languages. One cause is ownership bias: a model trusts its own answer more when that answer appears in the assistant role. This effect is well documented in English, but it is unclear whether it also appears in low-resource languages such as Banjarese and Dayak Ngaju. This study measures the extent of ownership bias and the calibration quality of four open-weight 8-bit LLMs when tested on school exam questions in Banjarese and Dayak Ngaju, drawn from the IndoMMLU benchmark. Four models (Qwen3-8B, Llama-3.1-8B, SahabatAI-Llama-8B, and SahabatAI-Gemma2-9B) are evaluated under two conditions: Own, in which the model’s answer is placed in the assistant role, and Reframed, in which the same answer is attributed to the user. We report accuracy, ownership bias, Expected Calibration Error (ECE), and Brier Score. Llama-3.1-8B shows the highest ownership bias (49.93 pp on Banjarese and 53.83 pp on Dayak Ngaju), while SahabatAI-Gemma2-9B shows almost no bias (0.91–1.33 pp) but assigns 100% confidence to every answer. The same pattern appears in both languages, indicating that ownership bias is driven by the model's architecture and training, not by the language being tested. AI assistants that are overconfident in English remain overconfident in Banjarese and Dayak Ngaju as well. Practitioners should treat verbalized confidence as an unreliable signal and use model-level calibration such as temperature scaling or logprob-based methods, rather than relying on prompt-level reframing alone.]]>
Copyrights © 2026
