白宫总统科技政策办公室主任 Michael Kratsios 在备忘录中表示，主要位于中国的外部行为体已对美国前沿 AI 系统展开有意的、工业规模（industrial-scale）蒸馏行动，美国政府计划对其采取打击，且此举发生在特朗普总统赴北京与习近平会晤的前数周。此指控又一次升温了有关先进 AI 研究被窃取的争议，Kratsios 指这些行动透过大量规避侦测的操作并具备协同行为，白宫将其定义为剥削美国创新，而非正常研究合作；中国驻华府大使馆则称该备忘录为「纯粹诽谤」。

文章聚焦于 AI「蒸馏」（distillation），即以大型模型输出训练较小模型。此技术在 DeepSeek 的案例中受到关注，指其在较低成本下据称推出更强产品。Anthropic 与 OpenAI 等美企表示，工业规模的蒸馏可在美方依赖先进晶片出口管制所形成的计算优势下缩小竞争差距，因为非法复制系统可在较低算力预算下重现核心能力。美国前美中关系研究所学者 Chris McGuire 认为，这削弱美国竞争地位并带来安全风险。

Kratsios 区分合法蒸馏与非法行为：前者可用于构建轻量化模型，但「未授权、工业规模蒸馏」是不可接受的。即使蒸馏模型常未达原模型效能，也因成本低廉而具商业吸引力。备忘录指出，美国各机构将共享未授权蒸馏情报并协同防御，并考虑对相关行为体采取制裁与问责；2025 年初 OpenAI 指出 DeepSeek 使用其 GPT 输出训练模型，至 2025 年 2 月 Anthropic 又将 DeepSeek、Moonshot 及 MiniMax 列为受指控对象。美国国会外交事务委员会亦推进法案，拟将从事蒸馏的实体列入出口黑名单 Entity List，以加强美国技术领先地位。

In a memo, White House Science and Technology Policy Office director Michael Kratsios said foreign actors—primarily entities based in China—have carried out deliberate, industrial-scale campaigns to distill U.S. frontier AI systems, and the U.S. government will move to crack down on this, including in the weeks before President Donald Trump’s meeting with Xi Jinping in Beijing. This claim has further intensified the dispute over alleged theft of advanced AI research; Kratsios said the operations use large-scale evasion techniques and coordinated behavior, and the White House defines them as exploitation of U.S. innovation rather than normal research collaboration, while the Chinese Embassy in Washington called the memo 'pure slander.'

The report centers on AI distillation—training smaller models from outputs of larger ones. The issue drew attention after DeepSeek was alleged to have used the method to launch a stronger product at lower cost. U.S. firms including Anthropic and OpenAI argue that industrial-scale distillation can narrow the advantage from which the United States benefits under advanced chip export controls, because illicitly copied systems can reproduce core capabilities with lower compute budgets. Expert Chris McGuire of the Council on Foreign Relations says this undermines U.S. competitiveness and creates security risks.

Kratsios distinguished legitimate distillation from abuse: the former can support lightweight model development, but unauthorised, industrial-scale distillation is unacceptable. Even if distilled models often trail originals in performance, they can still be commercially attractive because of much lower cost. The memo says U.S. agencies will share intelligence on unauthorised attempts and coordinate defenses, and officials are considering sanctions and accountability measures. OpenAI said in early 2025 that DeepSeek used its GPT outputs to train a model, and in February Anthropic named DeepSeek, Moonshot and MiniMax as target entities; the House Foreign Affairs Committee then advanced bills, including adding distillation actors to the export blacklist Entity List, to strengthen U.S. AI leadership.