The U.S. says China's AI progress is down to 'distillation.' But is it that clear cut?

U.S. AI companies and authorities are increasingly worried about a technique called distillation attacks, which threaten the effectiveness of Western frontier AI models. China and Russia are reportedly leveraging these attacks to create more capable AI models at lower costs. China has threatened to implement countermeasures if the U.S. uses these claims to restrict Chinese AI advancements.

Efforts to prevent distillation attacks have been underway since 2026, with major Western AI labs collaborating to address the issue. However, foreign actors continue to purchase logs from legitimate user conversations, complicating efforts to halt the practice. Distillation attacks involve training smaller language models using prompts and responses from larger, more advanced models.

Chinese AI developers have been accused of using this method to rapidly advance their models, such as Deepseek and Kimi K3, which rival those from U.S. companies like Anthropic and OpenAI in terms of intelligence. While distillation is generally considered legitimate, using it to train on other companies' models is seen as more malicious.

This practice raises questions about intellectual property and fair competition in AI development. The U.S. and China are set to meet on September 24, potentially leading to discussions about AI development and distillation attacks. However, stopping these attacks entirely may prove challenging, as detecting them can be difficult, and circumventing safeguards can be nearly impossible.