Multilingual and Cross-Cultural Challenges in LLM Refusal Detection: Evaluating Detector Reliability Across Languages

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The global deployment of large language models (LLMs) demands safety mechanisms that function reliably across linguistic and cultural contexts. Yet refusal detection—a cornerstone of LLM safety evaluation—remains predominantly designed, tested, and validated in English-centric settings.