Claims DeepSeek used "distillation" to extract model outputs
"Some extraction activities are linked to Russia"
OpenAI, the developer of ChatGPT, has raised suspicions about the Chinese artificial intelligence (AI) company DeepSeek, suggesting that DeepSeek is illicitly extracting outputs from U.S. AI models.
According to U.S. financial media outlet Bloomberg on February 12 (local time), OpenAI submitted a memo to the U.S. House Select Committee on the Strategic Competition between the United States and the Chinese Communist Party, warning that DeepSeek is extracting outputs from U.S.-made AI models.
OpenAI explained that DeepSeek used a technique called "distillation." Distillation is a method of creating a new model with similar capabilities by using the answers generated by another AI model as training data. By using this method, a lightweight model can achieve performance comparable to that of a large model. U.S. AI companies also use distillation to develop lightweight versions of their own models. For example, the lightweight version of Google's Gemini AI model, Gemini Flash, was developed by distilling the higher-tier model, Gemini Pro.
However, OpenAI argues that while distilling one’s own model to redevelop it into a lightweight version is acceptable, creating a distilled model by using someone else’s model without authorization effectively constitutes plagiarism. OpenAI also raised the issue that, during the distillation process, the safety mechanisms companies have implemented to prevent misuse in high-risk fields such as biology and chemistry could be neutralized.
According to Bloomberg, OpenAI stated in the memo, "We observed that accounts linked to DeepSeek employees developed methods to circumvent access restrictions and hid their sources to gain access to the models," adding, "The approaches used for unauthorized extraction are becoming increasingly sophisticated, and in some cases the extraction activities are linked to Russia."
On this basis, Representative John Moolenaar, the Republican chairman of the U.S. House Select Committee on the Strategic Competition between the United States and the Chinese Communist Party, denounced the situation, saying, "Steal, copy, and erase is the typical playbook of the Chinese Communist Party," and, "Chinese companies will continue to extract U.S. AI models and exploit them for their own gain."
This is not the first time allegations have been raised that Chinese companies are extracting U.S. AI. Previously, David Sacks, chair of the White House President’s Council of Advisors on Science and Technology, said in a media interview last year that there is evidence that DeepSeek illicitly extracted U.S. AI models.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
