Summary of Key Points
Recent months have seen an increase in mutual regulatory actions between China and the United States in the field of AI. For the first time, the U.S. has included AI models themselves within its export control measures. Anthropic has accused Alibaba of “stealing technology” while simultaneously covertly installing tracking code in its products. The Chinese Ministry of Industry and Information Technology (MIIT) has warned about the security risks associated with Claude, leading Alibaba to immediately ban all Anthropic-related products and considering classifying AI technology breaches as national security offenses. Both governments and companies are working together to establish a dual-layered barrier, turning the metaphorical “silicon curtain” in the AI industry into a reality supported by legal regulations and concrete measures.
1. The U.S. Takes a Revolutionary Step: AI Models Become Subject to Export Control
In the past, the U.S. only regulated the export of AI models through their weight files or source code. However, this time, AI models themselves (such as Anthropic’s Mythos/Fable) have been directly designated for control. On June 12, the U.S. Department of Commerce issued a letter to Anthropic, requiring it to obtain permission before exporting these models to countries like China and Russia. Even remote access by foreigners to the AI outputs generated in the U.S. or maintenance of AI infrastructure by foreign personnel could trigger licensing requirements.
Anthropic complied with the request, promising to proactively assess model security risks, cooperate with the government in reviewing future products, and report any malicious activities, which led to the lifting of the restrictions. This marks the establishment of a collaborative regulatory framework where U.S. companies will take the initiative to ensure compliance without waiting for government orders.
2. Anthropic’s Double Standard: Accusing Alibaba of Stealing Technology and Secretly Installing Trackers
On one hand, Anthropic wrote to U.S. lawmakers, claiming that Alibaba used 25,000 fake accounts and made 28 million requests to “distill” its models (in other words, copied its technology through extensive questioning). On the other hand, since March of this year, it has been secretly embedding tracking code in Claude, collecting users’ time zones, proxy servers, and even indicating any connections with Chinese AI companies—yet these details were not mentioned in any update notes.
This behavior is akin to a restaurant claiming perfect hygiene while using leftover food behind the scenes. It reflects a mutual dependence between U.S. governments and companies: governments rely on companies to gather technical information to maintain their technological advantage, while companies rely on government authorization to protect their own innovations.
3. Why Did Alibaba Ban Claude Faster than the MIIT?
While the MIIT only issued a safety warning (not a ban), Alibaba notified all its employees on July 3 that all Anthropic products (including the Claude model and related programming tools) must be removed by July 10, along with the cessation of API calls and related expenses. There are three main reasons for this:
1. Compliance to Avoid Penalties: Using these products after the warning would be seen as intentional non-compliance, which Alibaba wants to avoid becoming a regulatory target.
2. Responding to Accusations: Taking action to prove it didn’t steal technology is more convincing than issuing statements alone.
3. Proactive Planning: China is discussing classifying AI technology breaches as national security offenses, and since Alibaba’s Qianwen model is actively used internationally, it needs to secure its internal systems to prevent technology leaks.
4. The Dual-Barrier Is Closing: Is the “Silicon Curtain” Really Coming?
Both China and the U.S. are tightening restrictions on AI cross-border activities: the U.S. is limiting the export of models, and China is preventing technology breaches. Companies are actively cooperating with the governments in these efforts. The consequences include:
- First, the gradual elimination of unauthorized uses of products like Claude;
- Stricter regulations for Chinese models being exported overseas;
- Possible restrictions on currently open practices such as sharing model weights and using model distillation techniques.
Once all potential loopholes are closed, the “silicon curtain” in the AI industry will become a reality—cross-border flows of AI technology between China and the U.S. will be strictly controlled.
(Note: The term “distillation attack” refers to the process of using a large number of targeted questions to make an AI model produce similar outputs, thereby copying its capabilities; the “silicon curtain” is a metaphor for the isolation of AI technology across borders, similar to the Iron Curtain during the Cold War.)