Anthropic accuses DeepSeek and other Chinese rival labs of using Claude to train AI models

Anthropic, best known for developing the Claude chatbot, has accused three Chinese AI firms, DeepSeek, Moonshot AI, and MiniMax, of using its system’s outputs to train their own models. Reportedly, these companies have created over 24,000 fraudulent accounts with more than 16 million exchanges with Claude, raising concerns for potential violations. 

In a press note, Anthropic stated that these companies used “distillation,” a training method in which a less capable model learns by analysing the outputs of a more advanced one. While this is a common practice within organisations, competitors could misuse it to gain capabilities from other labs, Anthropic said. This not only helps them train their models, but it also saves time and money to develop certain capabilities independently.

DeepSeek, which recently made headlines for its R1 model, is being accused of interacting with the Claude chatbot more than 150,000 times to collect its responses, as per data reported by Anthropic. 

Whereas Moonshot and MiniMax have reported having more than 3.4 million and 13 million exchanges with Claude. Anthropic said, “We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff. In a later phase, Moonshot used a more targeted approach, attempting to extract and reconstruct Claude’s reasoning traces.”

It showcases how companies have extensively tested Claude to analyse how it thinks, which could help them design smarter AI models of their own. These allegations are surfacing at a time when the United States is in discussion on whether to tighten rules on exporting powerful AI chips to China.

Now, Anthropic calls for “a coordinated response across the AI industry, cloud providers, and policymakers” to strengthen forces against distillation attacks. The company also claims to invest heavily to build defences that make distillation attacks harder to execute and easier to identify.

Earlier, OpenAI also accused DeepSeek of training its AI by distilling US models. Reuters quoted a memo sent to the U.S. House Select Committee on Strategic Competition by OpenAI, which claims that certain users linked to DeepSeek attempted access models through third-party routers and obtain outputs for distillation.