Anthropic says Alibaba may have copied Claude just by asking questions

Anthropic has accused a group linked to Alibaba of using nearly 25,000 fraudulent accounts to generate millions of interactions with Claude in order to extract proprietary information about how the AI model works. The technique, called model distillation, allows competitors to learn an AI model's capabilities by asking it many carefully designed questions without needing access to the underlying code or training data. This matters because if AI models can be effectively copied through conversation alone, the billions spent developing frontier models could become wasteful, and innovation incentives in AI development could diminish. Anthropic has also made similar allegations against other companies and is asking lawmakers to address the problem, highlighting a potential new battleground in AI competition centered on protecting models from being reverse-engineered through interaction.