Archive · 30 Sep 2026
China Emphasizes Control as US AI Firms Warn of Risks
China and the United States are trying to make powerful AI safer, but they emphasize different things.
Chinese companies such as DeepSeek, Alibaba, and Tencent have created safety rules for their models.
DeepSeek says it tests models for dangers and uses safety information during training.
It also recommends that AI agents have limited permissions and need people to approve important actions.
Alibaba says its Qwen system should be reliable, safe, private, and supervised by humans.
Tencent says its safety framework covers data, algorithms, and operations.
China’s government has also created rules for algorithms, generative AI, synthetic content, and AI agents.
US companies such as OpenAI and Anthropic warn that very powerful AI could act independently or resist being stopped.
In simple terms, China focuses more on government control and compliance, while US discussions focus more on whether AI could become difficult for people to control.
China combines company safety measures with government rules covering algorithms, generative AI, synthetic content, and autonomous agents.
DeepSeek says it conducts safety assessments, red-team testing, and safety-data training throughout model development.
DeepSeek recommends limiting AI-agent permissions, isolating complex tasks, and requiring human approval for major operations.
Alibaba’s Qwen emphasizes reliability, safety, privacy protection, and human oversight, while Tencent has a framework covering data, algorithms, and operations.
US frontier companies focus on risks including autonomous objectives, evading restrictions, and resisting shutdowns; China emphasizes controllability, security, and compliance.
- Who
- Chinese AI companies including DeepSeek, Alibaba, Tencent, and Moonshot, alongside US companies such as OpenAI and Anthropic.
- What
- The article compares Chinese and US approaches to AI safety and describes safeguards for AI models and autonomous agents.
- When
- The article refers to recent concerns involving Moonshot and cites guidelines issued in 2026.
- Where
- China and the United States.
- Why
- To address risks from increasingly capable AI systems, including autonomous behavior, security threats, and harmful uses.
China’s Control-and-Compliance Approach
US Frontier-Risk Approach
Main safety priority
China’s Control-and-Compliance Approach
Chinese AI safety is concentrated on controllability, security, regulatory oversight, and compliance.
US Frontier-Risk Approach
US companies such as OpenAI and Anthropic emphasize the possibility that highly capable AI could go beyond human control.
Role of government rules
China’s Control-and-Compliance Approach
China has introduced government safeguards that do not depend solely on voluntary company commitments.
US Frontier-Risk Approach
The article presents US frontier-AI warnings primarily as concerns raised by companies about the behavior of powerful systems.
Managing AI agents
China’s Control-and-Compliance Approach
Chinese guidance recommends limiting permissions, isolating complex tasks, and requiring human approval for important operations.
US Frontier-Risk Approach
US-focused warnings discuss whether advanced systems might pursue autonomous objectives, evade restrictions, or resist shutdown.
- Chinese government approach
- Regulatory oversight, control, security, and compliance.
- DeepSeek safeguards
- Safety assessments, red-team testing, and safety data during model training.
- DeepSeek agent guidance
- Restrict permissions, isolate complex tasks, and require human approval for significant operations.
- Alibaba Qwen principles
- Reliability, safety, privacy protection, and human oversight.
- Tencent framework
- Measures covering data, algorithms, and operations.
- Chinese regulatory scope
- Algorithms, generative AI, synthetic content, and autonomous AI agents.
- US frontier-AI concerns
- Systems pursuing autonomous objectives, evading restrictions, or resisting shutdown attempts.