OpenAI said that Chinese DeepSeek stole ChatGPT data for AI training

Developer ChatGPT said he has evidence that Chinese company DeepSeek used OpenAI data to train its rival AI model. OpenAI’s rules state that users cannot copy the company’s services or use their results to build competing models.

This is a method of “distillation” – a process that allows smaller models to become more efficient by using the results of larger versions. Distillation is a common practice in the industry, but if DeepSeek used it to build a competitor, it violates OpenAI’s terms of service.

The release of the new DeepSeek model has shaken the AI ​​market, causing a collapse in the shares of key companies in the industry. Nvidia shares fell by more than 17%, resulting in a loss of almost $600 billion in one day. However, on Tuesday, the situation partially stabilized and Nvidia shares rose by 9%. Previously, entrepreneur David Sachs, who was responsible for the development of AI under the Trump administration, said that “data theft is quite possible.”

Experts believe that the theft of training data from companies such as OpenAI is common for smaller Chinese and American AI labs, since full-fledged training requires significant investment. DeepSeek said that it used a cluster of 2,000 Nvidia H800 graphics cards and spent $ 5.6 million to train its V3 model with 671 billion parameters, while training GPT-4 cost $ 100 million. Suspicions of data theft arose even during the launch of the Chinese model, when it claimed that “it is ChatGPT.”

OpenAI and its key partner Microsoft are reported to have investigated and blocked accounts that probably belong to DeepSeek and were used for distillation last fall. OpenAI is also currently facing copyright infringement allegations from publishers and content creators, including a lawsuit from The New York Times, which claims the company is training its models on their articles without permission.