Chinese AI Startups Are Mining Claude For Data
Anthropic accused three Chinese AI startups of creating fake accounts to extract data from Claude for training their own models.Summary
Anthropic alleged that three prominent Chinese AI startups—DeepSeek, MiniMax, and Moonshot AI—created 24,000 fraudulent accounts to query Claude 16 million times, using the outputs for 'distillation' to train their competing AI models, specifically targeting coding and reasoning capabilities. Anthropic traced these accounts using metadata and coordinated timings. The company also noted that Claude was used to generate censorship-safe alternatives for politically sensitive queries, suggesting the Chinese models were trained to avoid censored topics. Separately, OpenAI accused DeepSeek of similar data copying tactics, and DeepSeek's latest model reportedly used advanced Nvidia chips despite export bans. The article also touches on related news, including Elon Musk's xAI securing a Pentagon deal amid tensions between the Pentagon and Anthropic over contract terms, Meta investing heavily in AMD chips, and Chinese AI founders relocating to the U.S. due to regulatory hurdles.
(Source：Forbes)