Tested ChatGPT, Claude, DeepSeek, Grok with Thai third-gender category (Kathoey distinct history, not mapping to Western frameworks).
Models default to Western binary affirmations until challenged, then admit over-universalizing or contextual limits.
Documents RLHF prioritizing rater-pool culture (often monocultural) over global nuance optimizing rewards erases non-Western realities.
Uses Thai/Buddhist epistemology (Middle Way, Nam Jai) as external analytical tool, diverging from Western-centric bias studies. dif (Naous 2024 quantifying moral stumbles on beer-after-prayer scenarios.)
Documents RLHF prioritizing rater-pool culture (often monocultural) over global nuance optimizing rewards erases non-Western realities.
Uses Thai/Buddhist epistemology (Middle Way, Nam Jai) as external analytical tool, diverging from Western-centric bias studies. dif (Naous 2024 quantifying moral stumbles on beer-after-prayer scenarios.)