July 13, 2026, (Inside AI) — China’s Ministry of Industry and Information Technology (MIIT) has launched a project to build a national safety benchmark for generative AI models. The initiative targets six core dimensions: content safety, value alignment, robustness, fairness, privacy protection, and trustworthiness.
The MIIT’s research arm, the National Industrial Information Security Development Research Centre, is recruiting companies and experts to co-develop the benchmark. Applications were due Tuesday, according to a notice published Monday.
The benchmark will design a hybrid methodology covering 31 specific safety risks across five major categories. These include risks like data leaks, hallucinations, and biased outputs that current frameworks do not adequately address.
This move places China alongside the U.S. and Europe in a global push to regulate AI safety. The European Union’s AI Act and U.S. executive orders have set high bars, but China’s approach is distinct—it embeds state-led technical standards directly into industrial compliance.
The MIIT’s effort is not isolated. It builds on China’s earlier regulations, such as the 2023 generative AI rules requiring security assessments. The new benchmark aims to standardize testing, making it easier for companies to meet legal requirements.
“The institute said that current frameworks fail to meet complex safety-governance needs, requiring a standardised testing platform to support industrial compliance,” the notice stated.
Experts note that while global regulators share goals, methods diverge. The EU relies on risk-based legislation, the U.S. on voluntary commitments and agency actions, and China on mandatory technical benchmarks. This could fragment international AI governance.
Dr. Helen Toner, a former OpenAI board member, has argued that benchmarks must be dynamic. “Static tests quickly become obsolete as models evolve,” she said in a recent talk. China’s hybrid methodology may incorporate adaptive elements, but details remain sparse.
The 31 risks span technical and ethical domains. For example, robustness testing might involve adversarial attacks, while fairness checks could examine demographic biases. Privacy protection will likely assess data leakage in model outputs.
China’s tech giants—Baidu, Alibaba, and Tencent—are expected to participate. Their involvement could accelerate benchmark adoption but also raise concerns about self-regulation. Smaller firms may struggle with compliance costs.
Historically, China has used standards to shape industries. The 2017 AI development plan set ambitious goals, and this benchmark continues that tradition. It could become a de facto requirement for market access, mirroring how cybersecurity reviews work.
Internationally, the benchmark may influence global standards. China’s growing AI exports mean its safety norms could spread. However, critics worry about alignment with authoritarian values, such as censorship under “content safety.”
The U.S. has pushed back on Chinese tech influence, but on safety, there is room for cooperation. The UK’s AI Safety Summit saw initial dialogues, and benchmarks could be a common ground—if transparency is ensured.
Missing from the notice is how the benchmark will be enforced. Will it be voluntary or mandatory? How often will it be updated? These questions remain unanswered, leaving companies in limbo.
For now, the project signals Beijing’s intent to lead on AI governance. As models grow more powerful, the race for safety standards is as critical as the race for capabilities.
