AIR-Bench 2024: a safety benchmark based on regulatory and policy risk categories

The researchers introduce AIR-Bench 2024, which they describe as the first AI safety benchmark aligned with emerging government regulations and company policies. Based on the team's AI risk research AIR 2024, the benchmark decomposes 8 government regulations and 16 company policies into a four-tier safety taxonomy, with the finest tier containing 314 fine-grained risk categories, and builds 5,694 prompts accordingly, with manual filtering and manual review to ensure quality. The team used the benchmark to evaluate multiple leading language models to examine how well the models comply with these safety concerns, providing a foundation for evaluating model safety across regulatory jurisdictions and promoting the development of safer and more responsible AI systems.