China’s Ministry of Industry and Information Technology has launched a project to create a safety benchmark for evaluating AI models, reflecting increased attention to AI risks worldwide as regulators in the US and Europe enact stricter controls.
- MIIT's benchmark covers six key AI safety dimensions and 31 specific risks.
- Hybrid assessment combines automated and human evaluation to prevent AI misuse.
- Echoes global moves such as the US executive order and EU AI Act enforcement.
What happened
China’s Ministry of Industry and Information Technology has initiated the creation of a comprehensive safety benchmark to assess artificial intelligence models, particularly generative AI. The National Industrial Information Security Development Research Centre is spearheading this effort and calling on companies and experts to participate. The objective is to build a standardized testing platform capable of meeting complex industrial safety and compliance needs.
The benchmark will evaluate AI through six fundamental dimensions: content safety, value alignment, robustness, fairness, privacy protection, and trustworthiness. It will explicitly cover 31 specific safety risks across five categories using a hybrid methodology that integrates automated fuzzing and stress testing alongside human review, aiming to curb issues such as hallucinations, data leaks, and ‘jailbreak’ vulnerabilities.
Why it matters
This development highlights China’s strategic approach to AI governance as global authorities tighten control over emerging risks associated with large language models and generative AI. Existing frameworks have been deemed insufficient for the evolving complexities of AI safety, prompting the need for a more robust and technology-driven standard in a key global AI market.
Parallel global regulatory actions, including the US White House’s executive order for classified AI benchmarking and the European Union’s AI Act with significant compliance penalties, underscore heightened international concern over AI security, transparency, and ethical use. China’s initiative can position it to better manage the risks inherent in AI adoption while aligning with international standards.
What to watch next
The recruitment of industry stakeholders and AI experts to help develop the benchmark will be a critical phase, as broad participation could ensure the benchmark reflects diverse operational realities and technical challenges. The detailed design and implementation timeline of the hybrid testing methodology will be closely monitored by global observers and commercial AI developers.
Additionally, how this Chinese safety benchmark will interact with or influence international regulatory frameworks remains a key question. Observers will watch for potential alignment or divergence from the EU’s stringent data transparency rules and the US approach to classified AI capabilities, especially as the global AI ecosystem becomes increasingly interconnected.