AIR-Bench 2024: A Safety Benchmark Based on Risk Categories from Regulations and Policies

Regulations and policies reveal gaps in current safety testing: Existing AI safety benchmarks are often based on intuition or previous literature rather than actual government regulations and company policies, creating a misalignment between what researchers test and what regulators actually require.

A comprehensive taxonomy now exists for AI risks: AIR-Bench 2024 analyzed 8 government regulations and 16 company policies to create a structured framework with 314 specific risk categories, providing the first systematic mapping of how different jurisdictions and organizations define AI safety concerns.

Current language models show inconsistent safety alignment: When evaluated on AIR-Bench 2024's 5,694 test prompts, leading language models revealed varying levels of alignment with specified safety concerns, suggesting that developers may not be equally addressing all regulatory and policy requirements.

A practical tool for cross-jurisdictional AI assessment is now available: AIR-Bench 2024 enables companies and regulators to evaluate AI systems against multiple jurisdictions' requirements simultaneously, rather than conducting separate assessments for each region's rules.

Quality matters in safety benchmarking: The benchmark underwent manual curation and human auditing to ensure reliability, addressing a common weakness in AI safety testing where low-quality test cases can produce misleading results about model safety.