OpenAI suspects that China's DeepSeek AI models, significantly cheaper than Western counterparts, may have been trained using OpenAI's data. This revelation, coupled with DeepSeek's rapid rise in popularity, triggered a stock market plunge for major AI players. Nvidia, a key GPU provider for AI, suffered the largest single-day loss in Wall Street history, losing nearly $600 billion in market value. Other companies like Microsoft, Meta, Alphabet, and Dell also experienced significant drops.
DeepSeek's R1 model, based on the open-source DeepSeek-V3, boasts significantly lower training costs (estimated at $6 million) compared to Western models. While this claim is disputed, it fueled investor concerns about the massive investments in AI by American tech companies. DeepSeek's app rapidly climbed the US download charts, further highlighting the controversy.
OpenAI and Microsoft are investigating whether DeepSeek violated OpenAI's terms of service by using its API to "distill" data from OpenAI's models. OpenAI acknowledges that Chinese companies frequently attempt to replicate leading US AI models, emphasizing its efforts to protect its intellectual property and collaborate with the US government to counter such actions.
David Sacks, President Trump's AI advisor, supports the claim that DeepSeek used distillation, a technique that extracts data from larger models. He anticipates further measures from leading AI companies to prevent this practice.
The situation highlights the irony of OpenAI's accusations, given its own history of using copyrighted material to train ChatGPT. OpenAI previously argued that creating AI models like ChatGPT without copyrighted material is impossible, a stance supported by their submission to the UK's House of Lords. This position is further complicated by lawsuits from the New York Times and 17 authors alleging copyright infringement. The legal landscape surrounding AI training data remains highly contentious, particularly in light of a 2018 US Copyright Office ruling that AI-generated art is not copyrightable.
