On 11 February 2025, Singapore and Japan announced a Joint Testing Report, conducted under the purview of the AI Safety Institute (AISI) Network. This report is intended to evaluate the safety of Large Language Models (LLMs) across a variety of linguistic environments, with the aim of ascertaining the efficacy of AI safeguards in non-English languages. The testing process involved models in 10 languages across five harm categories, including privacy, crime, and intellectual property. This initiative is designed to address the limitations of AI safety that stem from its predominantly English-centric training, thereby contributing to the development of global evaluation standards for multilingual AI systems.
Original source