Latest / GNN Daily Breaking News / Chinese Military Researchers Tap U.S. AI Models to Train Defense Systems

Transcript

Chinese Military Researchers Tap U.S. AI Models to Train Defense Systems · GNN Daily Breaking News

• 0:00
  Speaker 1: A new investigation has revealed that Chinese military researchers are
• 0:04
  Speaker 1: systematically using outputs from US frontier AI models to train
• 0:08
  Speaker 1: their own defense systems, raising alarms in Washington about intellectual
• 0:13
  Speaker 1: property theft and national security risks. The practice, known as
• 0:18
  Speaker 1: model distillation, allows researchers to feed prompts into advanced American
• 0:22
  Speaker 1: AI systems such as those developed by OpenAI and Anthropic,
• 0:27
  Speaker 1: and then use the responses to train smaller specialized models.
• 0:32
  Speaker 1: By doing so, China's People's Liberation Army PLA can bypass
• 0:37
  Speaker 1: the enormous computing power required to build frontier models from scratch,
• 0:42
  Speaker 1: while still capturing their reasoning processes. A review of more
• 0:46
  Speaker 1: than eighty academic papers and patents linked to PLA institutions
• 0:50
  Speaker 1: shows widespread adoption of this technique. These distilled models are
• 0:54
  Speaker 1: being adapted for surveillance, cyber warfare, and tactical decision making,
• 1:00
  Speaker 1: giving China a shortcut to frontier capabilities without direct access
• 1:04
  Speaker 1: to restricted hardware or training data. US officials warned that
• 1:08
  Speaker 1: this undermines export controls designed to keep advanced AI out
• 1:13
  Speaker 1: of adversaries hands. The revelations come just weeks before high
• 1:17
  Speaker 1: level talks on AI governance and safety, where Washington is
• 1:21
  Speaker 1: expected to press Beijing on the issue. Critics argue that
• 1:25
  Speaker 1: the use of American models for military purposes highlights the
• 1:29
  Speaker 1: difficulty of enforcing digital export restrictions in an era where
• 1:33
  Speaker 1: access to cloud based systems is global. Chinese companies and
• 1:37
  Speaker 1: officials have pushed back AI startup Moonshot denied allegations that
• 1:42
  Speaker 1: its Kimi k III model relied on distillation from US systems,
• 1:46
  Speaker 1: insisting it was built on proprietary innovations. Beijing has accused
• 1:51
  Speaker 1: Washington of AI hegemonism, claiming the US is trying to
• 1:55
  Speaker 1: monopolize frontier technology while ignoring the fact that distillation is
• 2:01
  Speaker 1: a common practice worldwide. Strategically, analysts say the PLA's use
• 2:06
  Speaker 1: of distillation is not just about copying answers. It's about
• 2:10
  Speaker 1: capturing the reasoning frameworks of Western AI models, which could
• 2:14
  Speaker 1: enhance battlefield decision making and cyber operations. For the US,
• 2:19
  Speaker 1: the challenge is clear preventing adversaries from leveraging its breakthroughs,
• 2:23
  Speaker 1: even without direct access to advanced chips or training infrastructure. Ultimately,
• 2:29
  Speaker 1: this discovery underscores the fragility of export controls in the
• 2:33
  Speaker 1: digital age, while distillation is a legitimate research method. Its
• 2:38
  Speaker 1: use by military institutions for defense applications raises profound questions
• 2:43
  Speaker 1: about AI governance, intellectual property, and global security. Competition