Anthropic accuses Chinese companies of unfair practices

Contents

US-based AI startup Anthropic accuses three Chinese AI companies of setting up more than 24,000 fraudulent accounts with Anthropic's AI model Claude to improve their AI models. The companies DeepSeek, Moonshot AI, and MiniMax had generated more than 16 million interactions with Claude in this way to train their systems. Anthropic announced this on Monday in a blog post, citing violations of terms of use and regional access restrictions.

The Chinese competitors used a technique called “distillation,” where a less powerful model is trained based on the results of a more powerful model. Anthropic explained that distillation is a widespread and legitimate training method. For example, leading AI providers routinely distill their models to provide smaller and more cost-effective versions to their customers.

However, distillation can also be used for illegal purposes: Competitors can develop powerful competing products “in a fraction of the time and at a fraction of the cost required for independent development” using this method.

Similar pattern of behavior

The three Chinese companies proceeded according to a similar pattern, according to Anthropic. They used fraudulent accounts and proxy services to access Claude on a large scale and remain undetected. “The volume, structure, and focus of the requests differed from normal usage patterns and reflected intentional exploitation of features rather than legitimate use,” Anthropic writes. Using IP address correlations, request metadata, infrastructure indicators, and, in some cases, confirmation from partners who observed the same actors and behaviors on their platforms, they were able to “attribute each campaign to a specific lab with high confidence.”

The extent of the distillation activities of the various companies varied. According to Anthropic, DeepSeek conducted 150,000 interactions with Claude, while Moonshot and MiniMax recorded more than 3.4 million and more than 13 million interactions, respectively. Such campaigns are becoming “increasingly intense and sophisticated,” Anthropic complains, calling for “swift, coordinated action by industry players, policymakers, and the global AI community.”

OpenAI vs. DeepSeek

Numerous Chinese AI companies have recently unveiled their latest AI models, including Moonshot and MiniMax. According to the US daily Wall Street Journal, many of these new models have improved capabilities in logical reasoning and programming. The next-generation AI model from China's AI pioneer DeepSeek is expected to be unveiled soon. When the company presented its Reasoning Model R1 early last year, its competitive capability shocked the industry – especially since the company spent only a fraction of the immense costs of its US competitors for AI training. Doubts arose whether the AI infrastructure, costing hundreds of billions of US dollars, built primarily by US companies, is even necessary.

Soon, rumors and speculations, primarily from the US, suggested that DeepSeek had used distillation to develop its AI models. OpenAI accused DeepSeek of training its chatbot with ChatGPT. In September, the journal Nature published concrete information about DeepSeek's training for the first time. According to the report, the AI training cost less than 300,000 US dollars. The article on Deepseek-R1 shows how the Chinese company introduced an automated trial-and-error system, where the technique was rewarded when it identified correct answers.

Nevertheless, according to information from the Wall Street Journal, OpenAI sent a memo to members of the US House of Representatives earlier this month accusing DeepSeek of using distillation to imitate OpenAI's products.

Warnings from Anthropic

Anthropic now warns in the blog post that distillation attacks undermine export controls introduced to maintain the United States' leadership in AI over competitors like China. Distillation requires access to advanced chips on a large scale, Anthropic further states, leading the company to conclude, “Limited access to chips restricts both direct model training and the extent of illegal distillation.”

Furthermore, the actions of the Chinese companies pose “significant national security risks” to the United States. “Foreign labs distilling American models can then feed these unprotected features into military, intelligence, and surveillance systems.”

(akn)