| （英） |
The use of generative AI is rapidly expanding beyond a small group of specialists to a broad range of users, making its safety an increasingly important concern. In order to build a society in which people can use generative AI with confidence, it is essential to assess its safety appropriately and use the results to inform users' choices and guide system improvements. This talk focuses on Japanese-language large language models (LLMs) and discusses a range of safety risks associated with interactions with LLMs, including disinformation and misinformation, bias, discrimination, and outputs that violate public order and accepted moral standards. It then introduces efforts to evaluate safety and tune models based on the evaluation results. A variety of benchmarks have been developed and used to evaluate LLM safety. As part of these broader efforts, a collaborative project involving industry, government, and academia was launched in autumn 2025 to develop a safety benchmark, and work on the evaluation data is progressing steadily. Through safety evaluations, the project aims to provide useful information to both users and developers. For users, it seeks to provide information on LLM safety that is as accurate as possible, helping them make informed choices appropriate to their intended uses and circumstances. For developers, it provides insights that help identify areas where LLM safety needs improvement and guide concrete actions to address them. This talk introduces the project's activities and explores future directions for ensuring safety based on societal consensus. |