U.S. Companies Have Been Taking Measures To Prevent China To Take From Their AI Models. Now They’re Escalating.

Leading U.S. artificial intelligence companies are quietly deploying increasingly aggressive measures to prevent Chinese rivals from using their technology to accelerate their own development.

Anthropic recently admitted that it briefly embedded software inside its Claude Code programming assistant that silently checked whether users’ computers were configured with Chinese time zones or connected to internet domains associated with Chinese AI firms.

The monitoring tool was removed after software developers discovered it and privacy advocates criticized the company for effectively surveilling its own users. Anthropic described the feature as an experimental security measure that has since been rolled back in favor of other protections.

Companies such as Anthropic and OpenAI argue that Chinese firms are using a technique known as “distillation” to rapidly improve their own AI systems by learning from the responses generated by more advanced American models.

Distillation is a common AI training method that involves using a larger model as a teacher for a smaller one. While the process itself is legal and widely used across the industry, Anthropic and OpenAI argue that unauthorized large-scale use of their commercial systems violates their terms of service and amounts to intellectual property theft.

Anthropic has repeatedly claimed that Chinese AI companies have launched sophisticated campaigns involving thousands of fraudulent accounts to extract millions of responses from Claude. In a recent letter to U.S. senators, the company alleged that Alibaba’s Qwen AI team used roughly 25,000 fake accounts to generate more than 28.8 million interactions with Claude in an effort to improve its own technology.

The company previously made similar accusations against Chinese AI developers DeepSeek, Moonshot and MiniMax, arguing that such efforts have grown in both scale and sophistication. OpenAI has also warned U.S. lawmakers that Chinese competitors are attempting to replicate American AI systems through similar methods.

Chinese AI models continue narrowing the performance gap with American rivals. According to industry benchmarks cited by The Washington Post, Chinese developers have consistently matched the capabilities of leading U.S. models within months of their release.

Cybersecurity company Semgrep recently reported that a free coding model from Chinese developer Zhipu AI outperformed Anthropic’s Claude Opus 4.8 in identifying software vulnerabilities. The rapid progress has fueled concerns inside Washington that China’s AI industry could erase America’s technological advantage despite U.S. export controls on advanced semiconductors.

Anthropic has embraced that argument publicly, contending in a blog post earlier this year that preventing unauthorized distillation and tightening export restrictions could allow the United States to maintain a lead of 12 to 24 months over Chinese competitors.

A company spokesperson told the Post that these “attacks pose a serious threat to national security and undermine AI safety standards across the industry. That’s why we continue to speak openly about what we’re seeing and work closely with other labs, government, and partners on shared solutions.”

Those warnings have found support within the Trump administration and among several Republican lawmakers, who argue that preventing adversaries from gaining access to advanced AI technology is essential to maintaining U.S. leadership.

Not everyone agrees with that approach. Experts such as Irene Solaiman, chief policy officer at Hugging Face, told The Washington Post that portraying successful Chinese AI models as merely “stolen” risks underestimating China’s genuine engineering advances. Chinese companies have become increasingly efficient at building powerful, lower-cost AI systems, partly because years of U.S. export restrictions forced them to innovate with fewer computing resources.

That cost advantage is already attracting customers beyond China. Major startups, academic researchers, and even Fortune 500 companies have begun experimenting with Chinese open-source AI models because they are substantially cheaper than premium offerings from Anthropic and OpenAI, while still delivering competitive performance.

Meanwhile, blocking Chinese access remains difficult. Despite IP restrictions, government identification requirements and hundreds of thousands of banned accounts, Chinese developers continue accessing U.S. AI models through proxy networks, resellers, and third-party services.