US Administration Accuses Chinese Entities of Large-Scale AI Model Distillation, Announces Countermeasures

Introduction

The White House has issued a memorandum accusing foreign entities, mainly based in China, of conducting large-scale campaigns to extract capabilities from US advanced AI systems. The administration stated it will share intelligence with American AI companies, develop defensive best practices, and explore penalties.

Main Body

On April 23, Michael Kratsios, director of the White House Office of Science and Technology Policy, released a memo alleging that foreign actors—primarily in China—are engaged in deliberate, large-scale distillation campaigns. According to the memo, these actors use tens of thousands of fake accounts and methods to bypass security to systematically extract private information from US AI models, thereby taking advantage of US innovation. The administration indicated it would work with US AI firms to identify such activities, build defenses, and hold offenders responsible. Furthermore, the memo was released at the same time as a bipartisan bill in the House Foreign Affairs Committee that would create a process to identify foreign entities extracting key technical features from closed-source US AI models and impose sanctions. Representative Bill Huizenga (R-MI), the bill’s sponsor, described model extraction as a form of economic coercion and intellectual property theft. Chinese officials responded by rejecting the allegations. Liu Pengyu, a spokesperson for China’s embassy in Washington, stated that China opposes unfair treatment of its companies and emphasized its commitment to intellectual property protection through cooperation and healthy competition. Foreign Ministry spokesman Guo Jiakun described the US claims as false accusations and accused Washington of forcing technology, urging an end to suppression and a focus on mutual technological exchange. The accusations follow earlier statements by US AI companies. In February, Anthropic alleged that Chinese firms DeepSeek, Moonshot AI, and MiniMax had illegally extracted capabilities from its Claude chatbot using distillation. OpenAI, in a letter to lawmakers, accused DeepSeek of free-riding on its models. David Sacks, then serving as President Trump’s AI and crypto adviser, also claimed that DeepSeek had distilled knowledge from OpenAI’s models. Last year, DeepSeek released a large language model that competed with US counterparts at much lower cost, upsetting markets. Distillation is a known technique for training smaller, cheaper models from larger ones. Anthropic acknowledged its legitimate use but argued that unauthorized large-scale extraction allows competitors to gain powerful abilities in a fraction of the time and cost of independent development. The practice is not one-sided: San Francisco-based startup Anysphere recently stated that its latest product was based on an open-source model from Chinese company Moonshot AI. However, Kyle Chan, a fellow at the Brookings Institution, noted that distinguishing unauthorized distillation from legitimate data requests is very difficult, like "looking for needles in an enormous haystack." He suggested that information sharing and coordination among US AI labs could help, and that the federal government could assist with this. Chan also observed that the timing of the memo—before a planned mid-May state visit by President Trump to Beijing for a summit with President Xi Jinping—may affect the administration’s approach, as Trump might avoid increasing tensions before the meeting.

Conclusion

The US administration has formally accused Chinese entities of systematic AI model theft and announced plans to counter such activities, while China has denied the allegations and called for cooperation. The situation happens at a time when the performance gap between US and Chinese AI is narrowing, and a high-level diplomatic meeting is coming, leaving the future of enforcement and bilateral relations uncertain.

Vocabulary Learning

allegations
Claims or assertions that someone has done something wrong, often without proof.指控或斷言某人做了錯事,通常未經證實。
Example:Chinese officials responded by rejecting the allegations of large-scale AI model distillation.
bypass
To go around or avoid an obstacle, rule, or system.繞過或避開障礙、規則或系統。
Example:The memo claims that actors use methods to bypass security to extract private information from US AI models.
hold responsible
To consider someone or something accountable for an action or outcome.認為某人或某事應對某行動或結果負責。
Example:The US administration indicated it would work to identify activities and hold offenders responsible.
impose sanctions
To officially order penalties or restrictions against a country, organization, or individual.正式實施懲罰或限制措施針對某國、組織或個人。
Example:A bipartisan bill would create a process to identify foreign entities and impose sanctions for extracting technical features.
take advantage of
To use a situation or person for one's own benefit, often unfairly.利用某種情況或他人來獲取利益,通常是不公平地。
Example:The administration accused foreign entities of taking advantage of US innovation through systematic extraction.

Sentence Learning

The White House has issued a memorandum accusing foreign entities, mainly based in China, of conducting large-scale campaigns to extract capabilities from US advanced AI systems.
Present participle phrase 'accusing...' acts as a reduced relative clause describing the memorandum. The phrase 'mainly based in China' is a past participle phrase providing additional detail. This structure packs information concisely.現在分詞短語「accusing...」用作縮減關係從句,修飾備忘錄。短語「mainly based in China」是過去分詞短語,提供額外細節。這種結構能簡潔地包含信息。
Furthermore, the memo was released at the same time as a bipartisan bill in the House Foreign Affairs Committee that would create a process to identify foreign entities extracting key technical features from closed-source US AI models and impose sanctions.
Passive voice ('was released') shifts focus to the memo. The relative clause 'that would create...' defines the bill. 'Furthermore' is a linking adverb adding information.被動語態(「was released」)將焦點轉移到備忘錄上。關係從句「that would create...」界定該法案。「Furthermore」是連接副詞,用於補充信息。
According to the memo, these actors use tens of thousands of fake accounts and methods to bypass security to systematically extract private information from US AI models, thereby taking advantage of US innovation.
'According to' introduces the source. Infinitives 'to bypass' and 'to extract' express purpose. 'Thereby' is a linking word showing result, followed by a gerund phrase.「According to」引出信息來源。不定式「to bypass」和「to extract」表示目的。「Thereby」是表示結果的連接詞,後接動名詞短語。
However, Kyle Chan, a fellow at the Brookings Institution, noted that distinguishing unauthorized distillation from legitimate data requests is very difficult, like 'looking for needles in an enormous haystack.'
'However' signals contrast with previous information. The appositive 'a fellow...' identifies Kyle Chan. The noun clause 'that distinguishing... is difficult' acts as the object of 'noted'. The simile 'like...' adds clarity.「However」表示與前文對比。同位語「a fellow...」說明Kyle Chan的身份。名詞從句「that distinguishing... is difficult」作「noted」的賓語。明喻「like...」使表達更清晰。
The US administration has formally accused Chinese entities of systematic AI model theft and announced plans to counter such activities, while China has denied the allegations and called for cooperation.
'While' is a linking word showing contrast between two parallel actions. The sentence uses parallel verb phrases ('accused... and announced...'; 'denied... and called for...') for balanced structure.「While」是表示對比的連接詞,連接兩個並行動作。句子使用並行動詞短語(「accused... and announced...」;「denied... and called for...」)以達到平衡結構。