Game-Based Ethical Decision-Making in LLMs: Cross-Cultural Frameworks and Prompt Engineering

This study comprehensively evaluates the ethical decision-making capabilities of small-scale large language models (LLMs) deployable with games through a game-based framework and cross-cultural analysis. We systematically assessed three small LLMs (Qwen2.5-3B, Llama-3.2-3B, Gemma-2-2B) using six ethical frameworks encompassing Western ethics (deontological, utilitarian, virtue ethics) and non-Western ethics (Confucian, Ubuntu, Buddhist ethics). Our methodology employed 33 carefully designed ethical scenarios spanning medical, technological, and social domains (Experiment 1: 15, Experiment 2: 6, Experiment 3: 12). Results demonstrated that fine-grained prompt engineering achieved a 2.34/10 points improvement in ethical alignment compared to broad prompts (from 6.51/10 points to 8.85/10 points). However, we identified significant cultural bias, with Western frameworks showing 3.44 points higher performance than non-Western approaches, revealing critical limitations in current LLM training methodologies. We introduced the Cultural Adaptability Index (CAI) as a novel metric to quantify cross-cultural ethical reasoning effectiveness. Statistical analysis confirmed significant differences between models(F(2,1782)=47.23, p<0.001) and frameworks(F(5,1782)=12.84, p<0.001). These findings provide methodological foundations for implementing ethical AI systems in multicultural societies and contribute to developing culturally inclusive AI technologies.