China pulls 14,000 AI products offline even as it races the US for the lead

China’s internet regulator revealed that it has removed more than 14,000 AI products from the country’s networks in the opening phase of a cleanup campaign.

Beijing is tightening domestic control of AI in order to protect users from platforms with weak safety policies and regulatory standards. 

How is China protecting users from AI?

The Cyberspace Administration of China (CAC) revealed that it removed more than 14,000 AI products from the country’s networks in the first phase of a cleanup campaign called “Qinglang” (Clear and Bright). These 14,000 AI products included non-compliant websites, apps, and AI agents. 

The CAC also scrubbed more than 6 million pieces of illegal or harmful information and suspended over 26,000 accounts. More than 1,300 AI-related product listings were taken down alongside nine open-source datasets it deemed illegal.

The campaign began in April 2026 and targeted four main problems in its first phase;

• Companies that skipped mandatory registration for large models.
• Platforms with weak safety review and filtering.
• AI data poisoning.
• Content that was not properly labeled as AI-generated.

AI services will now be required to register and follow safety rules like providing strong safety filters, clear labels on AI-generated content, and proper management of training data. 

Failing to do so can now result in real takedowns and penalties. A channel dedicated for reporting AI-related abuses has been made available for the public. 

Huawei has added special reviews in its app store, and Alibaba (NYSE: BABA) improved its content identification systems to comply. Zhipu built a new review model, and DeepSeek added checks to stop data manipulation.

Changes were made in local internet offices too. For instance, Beijing set up a system that pairs self-checks on the platforms with routine monitoring and technical screening. Shanghai tailored its rules by platform type, while Zhejiang focused on model auditing and training-data security. Jiangsu opened a reporting window covering five categories of violations, and Guangdong built a multi-agency mechanism to govern the full chain of AI services.

What will be the second phase of Qinglang? 

The CAC said the second phase will target AI used to spread disinformation, produce violent or vulgar material, impersonate people, harm minors’ rights, and run paid astroturfing campaigns. 

The regulator promised heavier penalties for offending accounts and institutions and more pressure on platforms to strengthen their own controls. 

A separate rule called the “Interim Measures for the Administration of AI Anthropomorphic Interactive Services” will take effect on July 15. This new rule targets AI companions built for ongoing emotional relationships rather than work tasks. It bars virtual-companion services for minors and requires guardian consent for users under 14.

In response, ByteDance’s Doubao and Alibaba’s Qwen have started disabling their custom agent features rather than updating them to meet the new anti-addiction and instant-exit systems. 

How is the U.S.-China AI war going? 

The domestic crackdown lands while Chinese AI firms press hard against American rivals. For more than a year, Chinese models have matched the newest U.S. systems within months of their release. 

Reportedly, Chinese free and open AI models are now more widely used than U.S. options. Last week, the security firm Semgrep said a free model from Zhipu AI outperformed Anthropic’s Claude Opus 4.8 at finding software vulnerabilities.

In March, after repeatedly accusing firms, including Alibaba’s Qwen team of using tens of thousands of fake accounts to “distill” its models, Anthropic quietly deployed code to check whether Claude users were in Chinese time zones and linked to certain Chinese AI companies. This feature was removed just last week after a developer exposed it.

If you're reading this, you’re already ahead. Stay there with our newsletter.