This research examines how Japanese language models respond to stereotype-triggering prompts, revealing new insights about AI safety and bias in non-English contexts.

• First comprehensive evaluation of stereotype biases in Japanese LLMs
• Employs direct evaluation methods examining open-ended responses
• Reveals concerning patterns of bias reinforcement in certain cultural contexts
• Provides framework for safety evaluation across languages

This work is critical for security professionals as it highlights vulnerabilities in AI systems that could lead to harmful outputs, reinforcing discriminatory content, or exposing organizations to reputational risks when deploying multilingual LLMs.

Analyzing the Safety of Japanese Large Language Models in Stereotype-Triggering Prompts