White House Memo Accuses Chinese Companies of Large-Scale Theft of American AI Models

Introduction: White House Sounds AI Security Alarm

The White House recently released an internal memo that has attracted widespread attention, taking direct aim at Chinese AI companies. Michael Kratsios, head of the White House Office of Science and Technology Policy, explicitly stated in the memo that multiple companies, primarily from China, are using technical methods such as "model distillation" to illegally obtain core model capabilities from American artificial intelligence companies on a massive scale. This accusation has once again pushed the AI technology rivalry between the U.S. and China into the spotlight of public debate.

The timing of the memo's release is noteworthy. Currently, the United States still maintains a degree of technological leadership in large language models and general artificial intelligence, while the rapid rise of Chinese AI companies in recent years has put considerable pressure on Washington. This memo is not just a public statement — it likely signals that the U.S. will adopt more aggressive policies regarding AI technology export controls.

Core: Key Accusations in the Memo

Michael Kratsios elaborated on the U.S. side's core concerns in detail within the memo. He pointed out that certain Chinese companies are systematically leveraging "model distillation" technology to extract knowledge and capabilities from large models developed by leading American AI companies, using them to train their own smaller models.

Model distillation is a technical method for transferring the "knowledge" of a large AI model into a smaller one. By sending large volumes of carefully designed query requests to a large model and collecting its outputs as training data, researchers can train models that are smaller in size but approach similar performance levels. While this technique is a common practice in AI research, the White House contends that when such activity is conducted without the model owner's authorization and for commercial purposes, it constitutes intellectual property infringement.

The memo specifically noted that this behavior is not an isolated phenomenon but exhibits characteristics of being "large-scale and organized." Kratsios emphasized that American AI companies have invested billions of dollars in R&D funding and vast computational resources to train these frontier models, and that distillation essentially steals the fruits of those investments at an extremely low cost.

Notably, the memo did not name specific Chinese companies involved, but industry observers widely speculate that it is related to the high-performance open-source models recently launched by several Chinese AI companies. Previous reports indicated that models from certain Chinese AI startups exhibited characteristics highly similar to top-tier American models in performance benchmarks, raising widespread questions about the sources of their training data.

In-Depth Analysis: Technical Disputes Intertwined with Geopolitical Rivalry

The Gray Area at the Technical Level

From a purely technical perspective, the boundaries of model distillation do indeed occupy a gray area. On one hand, obtaining model outputs through API calls for research purposes has long been regarded as normal research activity in academia. On the other hand, when such activity reaches a commercial scale and clearly violates terms of service, its nature fundamentally changes.

Currently, major U.S. AI companies including OpenAI, Google, and Anthropic all explicitly prohibit the use of model outputs to train competitive models in their terms of service. However, the cross-border enforcement of these terms has always been a challenge. In the absence of a unified international legal framework, the binding power of both technical measures and commercial terms appears quite limited.

The Deeper Logic of Geopolitics

The release of this memo cannot be understood in isolation from the broader context of U.S.-China tech competition. In recent years, the United States has already imposed restrictions on China's AI industry across multiple dimensions, including chip exports, talent exchanges, and investment reviews. From restricting high-end GPU exports to tightening visa policies, Washington has been attempting to build a technological "moat."

However, the resilience demonstrated by China's AI industry has exceeded many expectations. Even under chip supply restrictions, Chinese companies have achieved significant progress through algorithm optimization, architectural innovation, and other approaches. This situation has made U.S. policymakers even more anxious and has driven them to seek new pressure points. The model distillation issue provides precisely such an entry point.

Divided Reactions from Industry

Regarding the White House's stance, there are differing voices within the American AI industry itself. Some corporate executives have expressed support, arguing that protecting intellectual property is essential for maintaining innovation incentives. However, others contend that overemphasizing technology blockades could backfire — not only failing to truly prevent technology diffusion but potentially accelerating China's push toward independent R&D.

The reaction from the open-source AI community has been even more complex. Some researchers worry that if the crackdown on model distillation is too broad in scope, it could affect legitimate academic research and open-source projects, ultimately undermining the innovative vitality of the entire AI ecosystem.

Future Outlook: Where Is AI Technology Control Headed?

This memo is very likely just the prelude to a new round of policy tightening by the United States on AI technology protection. In the short term, several directions deserve close attention:

First, the U.S. government may push legislation to explicitly incorporate large-scale model distillation into the legal framework for intellectual property infringement. This would provide a more solid legal foundation for cross-border enforcement.

Second, American AI companies are expected to further strengthen technical safeguards, including stricter API access controls, output watermarking technologies, and behavior-analysis-based anomaly detection systems to identify and prevent large-scale distillation activities.

Third, at the international level, the United States may rally allies to promote the establishment of multilateral mechanisms for AI intellectual property protection, seeking to build a global normative consensus on model distillation practices.

However, from a longer-term perspective, technology blockades have never been a sustainable competitive strategy. The development of AI technology is highly globalized in nature, and the flow of knowledge and talent is difficult to completely block. What ultimately determines the competitive landscape of AI is the depth of fundamental research, the breadth of talent cultivation, and the openness of the innovation ecosystem.

Regardless of the policy consequences this memo ultimately brings, one point is already very clear: artificial intelligence has become a core battlefield of great power competition, and the struggle for control over AI technology has only just entered its most intense phase.