DeepSeek, Temporarily Suspended in South Korea, Allegedly Transferred Personal Data Without Consent

Launched last January, DeepSeek R1 quickly shook up Silicon Valley and the AI ecosystem, including Nvidia, due to its performance and lower cost. The chatbot it powers was massively downloaded, but concerns about user data privacy were quickly raised, particularly in South Korea. 

By the end of January, due to uncertainties about the collection and use of user personal data, the Korean Personal Information Protection Commission (PIPC) requested explanations from DeepSeek, which admitted "some shortcomings" in its data management. The Commission decided the following month to remove the chatbot from local app stores, advising existing users to use it cautiously.

This temporary suspension aimed to thoroughly examine the company's practices and ensure they complied with the country's data protection laws. The South Korean regulator stated on April 24th that its investigation confirmed DeepSeek had indeed transferred personal information of South Korean users to companies in China and the United States without their prior consent. 

This data included content entered in AI prompts, as well as device, network, and application information. The PIPC recommended corrective measures, including the deletion of transferred data and the establishment of a legal framework for any future transfer of personal data abroad. The app's return to South Korean stores depends on effective compliance with these requirements.

The adoption of DeepSeek-R1 in China has been rapid, receiving strong support from the authorities. Beijing has actively encouraged its integration into key sectors such as justice, cybersecurity, and public administration.

The automotive industry is also on board: more than 20 Chinese manufacturers have expressed their intention to integrate DeepSeek-R1 into their future models. At the latest Shanghai Auto Show, BMW also announced its intention to incorporate DeepSeek AI models into the smart assistant of its vehicles sold in China.

While DeepSeek quietly launched last month DeepSeek-V3-0324, an update of its eponymous open source model DeepSeek-V3, rumors are rife about the upcoming availability of R2.

According to some sources, the startup, often suspected by its critics of bypassing U.S. restrictions to stockpile NVIDIA chips, has this time used local alternatives for training its new model: Huawei's Ascend 910B, which reportedly achieves about 91% of the efficiency of the A100 GPUs used for previous models.

This upcoming model is expected to have extended multilingual support, with better handling of Asian and European languages. Additionally, multimodal capabilities are mentioned, allowing the model to process text, images, audio, and video, paving the way for new applications in content creation and data analysis.

If these reports are confirmed, DeepSeek could once again disrupt the market balance...