Anthropic Accuses DeepSeek of Stealing Claude's Ability to Train Its Own Models

Illustration (photo: Unsplash)

JAKARTA - Anthropic accused three Chinese AI laboratories, DeepSeek, Moonshot AI, and MiniMax, of illegally using Claude AI to train their artificial intelligence models.

Called distillation, the action is usually intended to train a smaller and efficient model based on the output of a more sophisticated model.

However, if the technique is misused to copy the capabilities of a competitor's model at a cost and time that is much lower than independent development, the activity is considered to violate the terms of service.

In its official announcement, Anthropic said the three laboratories produced more than 16 million interactions with Claude through about 24 thousand fake accounts.

Anthropic highlighted that the illegal distillation model potentially does not have important safeguards. According to him, this model without protection can be used in military systems, intelligence, to mass surveillance.

The risk is even greater if the model that has been distilled is then released as open-source, so that its capabilities spread without control.

DeepSeek has recorded more than 150 thousand interactions, with a focus on the ability to reason and create alternative answers that are "sensor-safe".

Meanwhile, Moonshot AI, reportedly has conducted more than 3.4 million interactions, targeting agent-based reasoning, tool use, coding, and computer vision.

Meanwhile, the MiniMax laboratory was recorded as the largest, with more than 13 million interactions, especially in the ability to code and orchestrate tools.

In response, Anthropic strengthened its classifier-based detection system and behavioral fingerprinting to identify distillation patterns, including efforts to extract chain-of-thought reasoning.

The company is also improving verification of educational accounts, security research programs, and startup organizations - a path often used to create fake accounts.

In addition, Anthropic stated that it was developing security at the product, API, and model levels to reduce the effectiveness of model outputs for illegal distillation purposes, without disrupting the legitimate user experience.