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

Japanese LLMs show concerning safety gaps: The study found that LLM-jp, a Japanese-native model, had the lowest refusal rate when presented with stereotype-triggering prompts and was more likely to generate toxic and negative responses compared to models trained on English or Chinese, suggesting safety issues specific to Japanese language models.

Direct evaluation methods reveal hidden biases: By using open-ended prompts rather than asking models to choose between pre-written options, researchers uncovered stereotypical responses that indirect evaluation methods might have missed—highlighting the importance of how we test AI systems for bias.

Prompt wording matters significantly: The research demonstrated that how a question is phrased in Japanese substantially influences whether models generate biased or harmful content, meaning the same underlying bias can be triggered or suppressed depending on specific word choices.

Non-English AI safety research is critically needed: This study addresses a major gap in AI safety research, which has historically focused on English-language models. As Japanese LLMs are increasingly adopted, understanding their specific safety vulnerabilities is essential for responsible deployment.

Social group categories reveal systematic patterns: The researchers tested 301 different social group terms across age, gender, and other attributes, suggesting that stereotypes in Japanese LLMs may follow predictable patterns that could inform targeted safety improvements.