China’s AI catch-up complicates the global safety debate

China’s AI sector is narrowing the capability gap with the United States at the same time that both countries are arguing more intensely about safety, security and control. An Associated Press report published on September 16 says Chinese companies including Moonshot AI, DeepSeek, Z.ai and Alibaba have released models that increasingly challenge the lead of U.S. labs such as OpenAI and Anthropic. The U.S. still leads in many areas, including research influence and the top tier of model capability, but recent benchmarks and expert assessments suggest the gap is no longer comfortable or stable.

The competitive picture is changing quickly. Moonshot AI’s Kimi K3 drew attention in July with performance close to leading U.S. models and briefly ranked near the top of global model-intelligence tables. DeepSeek’s R1 had already changed the market in early 2025 by showing that Chinese models could compete at lower cost, and the company later previewed V4 with improvements in knowledge, reasoning and agentic capability. Z.ai released GLM-5.2 in June and GLM-5.3 in August, while Alibaba previewed Qwen3.8-Max in July as its strongest model in that line. None of these releases alone proves a permanent shift, but together they show a more resilient Chinese model ecosystem than chip restrictions were meant to allow.

That resilience is forcing a broader policy question. For years, U.S. export controls focused on limiting China’s access to advanced chips and chipmaking tools. Those controls still matter, especially for training the largest frontier systems, but they do not fully stop model progress. Chinese labs can optimize architectures, rely on open-weight research, train around hardware constraints and learn from the behavior of powerful foreign systems. AP cites experts who argue that the U.S. may now need to think not only about inputs such as chips, but about outputs such as model behavior and capability transfer.

Washington has already accused Chinese AI developers of industrial-scale distillation, a process in which one model is trained to imitate the outputs or capabilities of a stronger system. U.S. agencies including the FBI, NSA and CISA recently alleged that Chinese developers extracted capabilities from advanced American models through targeted distillation activity. Anthropic has also accused Chinese startups of routing some user requests to Claude. Beijing rejects the U.S. claims, saying distillation is widely used across the industry and accusing Washington of trying to monopolize AI. The dispute underscores how difficult it is to separate technical imitation, ordinary competitive learning and national-security concerns.

China is also warning about AI risks in its own terms. Chen Yixin, head of China’s state security ministry, has described threats ranging from deepfake-enabled political rumors to cyberattacks and data leaks. Chinese leaders continue to call for faster AI development and greater technological self-reliance, but they also say risks must be controlled and global governance should be discussed. That creates a complicated picture: the country is racing to catch up with U.S. frontier labs while also building its own language around security, ideology and governance.

The timing matters because U.S.-China relations are entering another high-stakes moment, with Chinese President Xi Jinping expected to visit Washington this month. AI will not be the only issue on the table, but it increasingly shapes trade policy, cybersecurity, military planning, industrial strategy and global technology standards. If the U.S. and China treat AI solely as a winner-take-all race, coordination on safety will be difficult. If they acknowledge shared risks, they may still disagree over verification, openness and control. The AP report shows why the global AI safety debate cannot be understood only through Silicon Valley. It is now inseparable from geopolitics.