Semantic Amplification of Color Bias in Large Language Models: Evidence from Cross-Linguistic Discrete Choice Experiments

Large language models (LLMs) are increasingly deployed in decision-making contexts, yet systematic biases in discrete choice remain poorly understood. We conducted 4,000 color selection trials across 10 state-of-the-art models (5 US-origin, 5 China-origin) to investigate how representational format, model origin, and prompt language influence choice behavior. In Study 1 (N=2,000), models selected from hex color codes, revealing a striking 67.2% preference for blue, nearly seven times the expected uniform rate. Study 2 (N=2,000) replicated the design using natural language color names instead of hex codes. Contrary to the hypothesis that blue preference stems from hex code artifacts in training data, natural language amplified the bias: blue selections increased to 91.8% (+24.6 percentage points), with choice diversity collapsing 70% (Shannon entropy: 1.720 → 0.506 bits at T=0.0). Six of ten colors received zero selections with names, compared to all colors being chosen with hex codes. Position bias was robust across both representations (position 1: 14.1% selection rate, χ2=77.23, p<0.001). Cross-cultural effects were minimal: Chinese models prompted in Chinese did not prefer red (3.85% overall), and origin effects after FDR correction were small (Cramér’s V≤0.102). Temperature robustness tests (T=0.0) confirmed biases persist under deterministic sampling. These findings demonstrate that LLM choice biases are semantically mediated rather than representational artifacts, suggesting learned linguistic associations, potentially from ecological valence patterns in training corpora, drive systematic preferences. Our results show that model origin does not produce culturally differentiated color preferences and highlight the need for representation-aware bias evaluation protocols in LLM deployment.