A State Department diplomatic cable sent Friday to US embassies and consulates worldwide has instructed American diplomats to warn host governments that Chinese AI companies, led by DeepSeek, are conducting coordinated industrial-scale campaigns to extract proprietary capabilities from American AI models, escalating the US-China AI conflict from a trade dispute into a formal intelligence matter.
The cable, reviewed by Reuters and not previously made public, instructs diplomatic staff to discuss concerns about "extraction distillation of US AI models" with international counterparts and to "alert about the dangers of employing AI models derived from US proprietary AI frameworks." A separate demarche has been sent to Beijing directly. The State Department's action follows a White House memo issued Thursday by Office of Science and Technology Policy Director Michael Kratsios, which accused Chinese entities of systematically targeting American frontier AI labs. The coordination between the White House memo and the State Department cable indicates this is a structured escalation, not an isolated complaint, and that the Trump administration intends to internationalize the issue by recruiting allied governments into a shared response framework.
The specific allegations are more granular than the diplomatic language suggests. Anthropic disclosed in February that DeepSeek, Moonshot AI, and MiniMax generated more than 16 million interactions with its Claude models in what the company characterized as a systematic attempt to extract model capabilities through volume-based distillation. OpenAI repeated to the House Select Committee on China in February that DeepSeek had targeted its systems to imitate ChatGPT's outputs and use them for training. Independent researchers alleged in June 2025 that an updated version of DeepSeek's R1 model showed evidence of distillation from Google's Gemini. The pattern across multiple US frontier labs, documented independently and now being shared through the Frontier Model Forum under an April 6 agreement between OpenAI, Anthropic, and Alphabet, forms the evidentiary basis for the State Department's global campaign.
The technical concept at the center of the allegation is knowledge distillation, a legitimate and widely used machine learning technique in which a smaller, cheaper model is trained using the outputs of a larger, more expensive one. The practice is not inherently illicit. OpenAI itself used distillation to train smaller versions of GPT-4. The specific allegation against DeepSeek is not that distillation was used as a technique but that it was applied to proprietary commercial models without authorization, at scale, using tens of thousands of proxy accounts and jailbreaking methods to circumvent the usage restrictions those companies had implemented.
The State Department cable acknowledged the nuance directly: "AI models created from covert, unauthorized distillation efforts allow foreign entities to launch products that seem to perform comparably on specific benchmarks at a significantly lower cost but fail to replicate the complete capabilities of the original system." That last clause matters. The allegation is not that DeepSeek stole a perfect copy of GPT-5. It is that unauthorized distillation gave Chinese firms a substantial head start, reducing training costs and development time while also, the cable alleged, stripping safety protocols and alignment features from the resulting models.
China's response has been consistent and dismissive. Foreign Ministry spokesperson Guo Jiakun called the White House's Thursday accusations "groundless" and characterized them as "deliberate attacks on China's development and progress in the AI industry." The Chinese Embassy in Washington did not respond to Friday's Reuters request for comment. Beijing's position is that it places significant emphasis on protecting intellectual property and that Washington's claims represent geopolitical containment dressed as IP enforcement. That rebuttal will not satisfy the bipartisan House Foreign Affairs Committee, which this week unanimously advanced legislation to create a sanctions framework targeting foreign entities that extract "key technical features" from closed-source US AI models.
The Strategic Calculation Behind Going Global
The decision to send a worldwide diplomatic cable rather than simply issuing a domestic policy statement reflects a specific strategic calculation. The US cannot contain Chinese AI development through chip export controls alone. The October 2022 semiconductor restrictions and subsequent tightening have slowed but not stopped Chinese frontier model development. DeepSeek V4, released Thursday at near-frontier capability and a fraction of US model pricing, is the most visible evidence that the hardware bottleneck has not been decisive. If the most effective path to Chinese AI advancement has shifted from acquiring restricted hardware to extracting capabilities from American software through distillation attacks, then the enforcement response has to shift accordingly.
Internationalizing the warning serves two purposes. It pressures allied governments to restrict their own institutions and companies from adopting AI models that may contain illicitly distilled US technology, a market access issue as much as a security one. And it builds the international coalition that any credible enforcement mechanism against Chinese AI firms will require. The Frontier Model Forum's information-sharing arrangement between OpenAI, Anthropic, and Alphabet represents the private sector dimension. The State Department cable represents the diplomatic dimension. The House legislation under consideration represents the statutory dimension. The US government is assembling all three simultaneously, and Friday's cable is the clearest signal yet that the assembly is complete.
Also read: Utah's medical board just called for the suspension of America's first AI prescription program and the industry should take note • GPT Image 2 generating Gemini watermarks exposes how AI-generated content has contaminated the training data ecosystem • DeepSeek V4 Pro costs 15x more to run than V3.2 and the efficiency narrative just got complicated
