Are China’s ‘AI tigers’ cheating? US rival Anthropic alleges some are

Anthropic, a leading AI company based in the United States, has leveled serious allegations against three prominent Chinese AI laboratories. The firm claims that these labs have unlawfully extracted capabilities from its Claude model to enhance their own systems, a situation that raises significant national security alarms. The accused Chinese firms—DeepSeek, Minimax, and Moonshot AI—are alleged to have created more than 24,000 fake accounts to train their models. According to Anthropic's recent blog post, they purportedly utilized over 16 million interactions with the Claude model through a process known as distillation. This practice, common in the AI sector, typically involves training models to create more cost-effective versions, but most leading AI providers, including Anthropic, explicitly prohibit such actions. Notably, Claude is not available for use in China. This revelation follows similar claims made by Anthropic's competitor, OpenAI, which earlier this month accused DeepSeek and other Chinese companies of illegally distilling its ChatGPT models over the past year. In a memo to the US House Select Committee on China, OpenAI expressed concern over DeepSeek's rapid advancements, suggesting they stem from unauthorized use of capabilities developed by American AI firms. DeepSeek made headlines last year by launching a highly competitive model that nearly rivaled industry leaders like ChatGPT, despite requiring less computational power. This development sparked discussions regarding the effectiveness of US technology and export controls, with OpenAI indicating it is investigating whether DeepSeek improperly used its models. Anthropic has raised alarms about the potential dangers of these illicitly distilled models, warning that they might lack essential safety features that reputable companies implement. Such models could pose risks if utilized for cybercrime or bioweapons, and they might empower authoritarian regimes to execute offensive cyber operations and engage in disinformation campaigns. As China’s AI sector continues to surge forward, questions arise about whether US export controls are effective. Anthropic argues that the successful development of high-performance models by these Chinese labs through distillation underscores the need for such restrictions. The company maintains that cutting-edge AI advancements cannot thrive solely on innovation without access to advanced technology. Alongside DeepSeek, Minimax and Moonshot AI's Kimi model have gained prominence, earning the nickname “AI tigers.” All three labs currently rank among the top 15 models on the influential Artificial Analysis leaderboard, highlighting their growing influence in the AI landscape.