Updated
Updated · Gizmodo · Aug 4
Anthropic Accuses 4 Chinese Labs of Illicitly Distilling Claude as US-China AI Tensions Rise
Updated
Updated · Gizmodo · Aug 4

Anthropic Accuses 4 Chinese Labs of Illicitly Distilling Claude as US-China AI Tensions Rise

3 articles · Updated · Gizmodo · Aug 4

Summary

  • Anthropic said DeepSeek, Moonshot, MiniMax and Alibaba illicitly used Claude to train their own AI models, widening a dispute over whether Chinese labs are copying U.S. frontier systems.
  • Model distillation lets a smaller “student” model learn from a larger “teacher” model’s outputs and reasoning, sharply cutting the compute and cost needed to approach advanced performance.
  • The company framed the issue as misuse rather than clearly illegal conduct, reflecting still-unclear rules around distillation even as Washington promises a tougher line on Chinese access to U.S. AI.
  • The fight has intensified since DeepSeek’s R1 launch in early 2025, when OpenAI similarly accused the startup of distilling ChatGPT after R1 neared frontier benchmarks at far lower cost.
  • Beyond geopolitics, critics warn distilled models may lose safety guardrails while inheriting biases, allowing risky capabilities to spread faster and more cheaply.

Insights

When smaller AI systems learn advanced capabilities but forget safety guardrails, what happens when these compressed models reach hackers?
Are claims of AI intellectual property theft just a smokescreen for tech giants trying to outlaw methods that democratize artificial intelligence?
If watermarks fail to prove AI copying, how will tech giants stop rivals from secretly extracting their multi-million dollar secrets?