OpenAI has claimed that the Chinese company DeepSeek used advanced methods to extract results from leading American AI models to train its own chatbot, R1.

The document alleges that DeepSeek used so-called distillation techniques an approach in which one AI model uses the responses of another to train and replicate similar capabilities. According to OpenAI, the Chinese company developed "disguised methods" to bypass safeguards designed to prevent misuse of the technology.

OpenAI launched a joint investigation with Microsoft into possible violations shortly after the release of the R1 model. The companies are investigating whether the data could have been obtained unauthorizedly. OpenAI emphasizes that such practices are becoming increasingly sophisticated, despite current restrictions and terms of use.

The company also cited commercial and geopolitical risks. Unlike American developers, DeepSeek and other Chinese players do not charge a subscription for access to their models, creating competitive pressure on OpenAI and Anthropic, which have invested billions of dollars in infrastructure and monetized premium features.

Furthermore, OpenAI raised national security concerns, stating that DeepSeek's model limits responses to topics sensitive to Chinese authorities. The company believes that distillation can undermine security mechanisms, increasing the risk of abuse, including in fields such as biology and chemistry.