White House Technology Director: Will Combat Industrialized Theft of US AI
Michael Kratsios, head of the White House Office of Science and Technology Policy, warned in a memorandum that foreign entities, primarily based in China, are conducting "industrial-scale" model distillation operations to steal the capabilities of US cutting-edge AI systems, and pledged to take measures to protect US AI innovation. The memorandum noted that these entities are systematically calling and distilling US models through thousands of proxy accounts, replicating some performance at a cost far lower than independent research and development, while also removing safety and neutrality constraints.
Several English media outlets cited the memorandum stating that this type of "industrial distillation" has escalated to a national security and technological sovereignty issue. The US government is requiring various departments and AI companies to share intelligence to identify and block suspicious access and data flows. Previously, Anthropic and OpenAI publicly accused several laboratories based in China of circumventing their model capabilities through distillation, making "whether distillation constitutes intellectual property and security risks" one of the core controversies in the global AI industry.
Source: Public Information
ABAB AI Insight
This statement elevates "model distillation" from an academic and engineering tool to a "technological sovereignty conflict tool." Past intellectual property disputes were mainly focused on chips, source code, and physical devices, but now the core assets have become parameter weights and behavioral distributions. Techniques like distillation essentially compress the external behavior of a black-box model into a new model, making it possible to replicate capabilities without obtaining the weights, which structurally undermines the effectiveness of traditional IP protection measures.
From a global technology landscape perspective, the US characterizing this as "industrial theft" marks AI as a strategic technology akin to advanced processes, satellites, and encrypted communications, entering a framework of "export controls + opponent non-proliferation." Once distillation is included in regulatory and enforcement narratives, there may be more granular controls in the future regarding interface call frequency, model openness levels, cross-border API access, and cloud deployment locations, essentially limiting capability spillover by controlling the inference process.
This conflict also exposes the narrative tension within the AI industry: the US advocates for large-scale data scraping training as fair use domestically, while externally accusing other countries of "free-riding distillation," which can easily be questioned as a "double standard" in international discourse. However, from a power structure perspective, the party at the technological forefront has a natural incentive to distinguish between "disputes over underlying training data" and "replication of upper-level capabilities," with the former being vague and the latter being publicly defined to maintain technological advantage and pricing power.
At a deeper level, this represents a shift from "hardware blockade" to "behavior blockade." Chip embargoes address computing power issues, while distillation disputes target the path of "replicating capabilities using existing public interfaces"; the former uses physical control, while the latter controls APIs and behavioral space through legal and contractual means. As enforcement and cases surrounding distillation gradually increase, the global AI ecosystem will be forced to rebalance between "open interfaces promoting innovation" and "tightening interfaces to prevent capability outflow," potentially exacerbating the divide between open AI and nationalized AI camps.