Combatting Chinese Distillation Attacks: The White House's Response to AI Theft
Sign up for the Artificial Intelligence myFT Digest to receive the latest updates directly to your email inbox. The White House has accused China of engaging in large-scale theft of intellectual property from American artificial intelligence labs. The US government has evidence that foreign entities, primarily based in China, are conducting deliberate campaigns to extract cutting-edge AI systems. This accusation comes amidst rising tensions over Chinese groups allegedly stealing advanced American AI research, as both countries compete to lead in the AI technology sector. The White House plans to crack down on this practice to protect US innovation.
China's DeepSeek has been accused of using distillation, a process of training smaller AI models based on larger ones' output, to develop a powerful product at a lower cost. The White House aims to share information with American AI companies about unauthorized distillation attempts by foreign actors and help them coordinate against such attacks. Chinese campaigns are reportedly using proxy accounts and jailbreaking techniques to evade detection and expose proprietary information. The US government is exploring measures to hold foreign actors accountable for these industrial-scale distillation campaigns.
Technology security expert Chris McGuire highlighted that Chinese AI firms are using distillation attacks to replicate the core capabilities of US models due to deficits in AI computing power. He suggested banning Chinese groups from accessing US models, sanctioning entities involved in distillation, and tightening export controls on AI chips to prevent unauthorized access. US AI companies, including Anthropic and OpenAI, have raised concerns about Chinese distillation attacks, which they believe undermine the competitive advantage of US models.
While distillation is a legitimate part of the AI ecosystem when used appropriately to create lighter-weight models, the White House deems industrial distillation aimed at undermining American research and development as unacceptable. Models created through unauthorized distillation campaigns may not match the original models' performance but can benefit foreign groups due to lower costs. American AI companies fear that distilled models pose national security risks as they lack safeguards against potential threats like bioweapons or cyber attacks.
In February, Anthropic accused three Chinese AI companies, including DeepSeek, of distillation attacks on its models. OpenAI also reported evidence of DeepSeek using outputs from its models to train its own model in violation of terms of service. Concerns about national security risks have prompted the House Foreign Affairs Committee to pass bills aimed at impeding China's progress in the AI race. One bill focuses on addressing distillation by potentially adding groups engaging in it to the export blacklist, making it challenging for US companies to sell technology to them.