China leads first global standard for automotive AI safety
China spearheads the first international standard for automotive AI safety, introducing measurable failure-rate limits and lifecycle risk controls for autonomous driving.
China has taken the lead in drafting the world's first international standard for automotive AI safety, a project officially approved by the Standardization Administration of China (SAC). This marks the first time Beijing has headed a global standard in this domain.
The initiative is co-led by experts from China and the United States, with contributions from specialists in Germany, France, Japan, and other nations. The standard targets intelligent connected vehicles (ICVs), where AI acts as the 'digital brain' governing operational safety.
Existing global guidelines on automotive AI safety remain broad and principle-based, lacking specific implementation rules for ICVs and autonomous driving scenarios. The new standard aims to fill that gap by translating general principles into measurable benchmarks and practical safeguards.
Li Bo, a senior researcher at the China Auto Standardization Research Institute, explained that the standard establishes quantitative safety indicators and comprehensive risk controls spanning safety analysis, testing validation, and operational monitoring. A key metric is the failure rate for autonomous driving: no more than one incident per 10,000 hours of operation. This target is further broken down and allocated to AI subsystems across various driving scenarios.
The standard also addresses challenges posed by AI large models, such as limited interpretability and inherent uncertainty. It analyzes failure modes and hazard mechanisms across the full lifecycle, converting them into concrete requirements for AI system architecture, model design, hardware security, and dataset development. Additionally, it introduces layered AI diagnostics and multi-dimensional redundancy designs to create robust safety defenses.
Li noted that the standard provides risk management guidelines for the entire lifecycle of AI safety in ICVs—from design, testing, and development to operational monitoring. It systematically tackles safety hazards arising from systemic failures or insufficient AI capability, which may stem from models, data, computing power, algorithms, and model iteration processes.