ChatGPT Maker Suspects China’s Dirt Cheap DeepSeek AI Models Were Built Using OpenAI Data — and the Irony Is Not Lost on the Internet

OpenAI suspects that China's DeepSeek AI models, significantly cheaper than Western counterparts, may have been trained using OpenAI data. This revelation, following Nvidia's massive stock drop, has sparked concern within the US tech industry and prompted President Trump to call it a "wake-up call."

DeepSeek's R1 model, built upon the open-source DeepSeek-V3, boasts significantly lower training costs (estimated at $6 million) and computational requirements compared to Western models like ChatGPT. While this claim is disputed by some, it has fueled investor anxieties about the billions being invested in AI by American tech giants. DeepSeek's app also rapidly climbed US download charts, further highlighting its impact.

OpenAI and Microsoft are now investigating whether DeepSeek violated OpenAI's terms of service by employing "distillation," a technique that extracts data from larger models for training. OpenAI confirmed to Bloomberg that it actively combats such practices, emphasizing the importance of collaboration with the US government to protect its technology. David Sacks, President Trump's AI czar, corroborated these concerns, citing substantial evidence of knowledge distillation from OpenAI models.

The situation is ironic, given OpenAI itself faces accusations of using copyrighted internet content to train ChatGPT. This hypocrisy has been widely noted, with critics pointing out OpenAI's previous justification that using copyrighted material is necessary for training leading AI models. OpenAI's stance is further complicated by the ongoing lawsuits from the New York Times and 17 authors alleging copyright infringement. These lawsuits highlight the contentious issue of training AI models on copyrighted material, a key debate within the rapidly evolving generative AI landscape. The legal precedent is further muddied by a 2018 US Copyright Office ruling that AI-generated art cannot be copyrighted.