Microsoft, a primary investor in OpenAI, is now exploring whether the Chinese company DeepSeek used nefarious methods to train its reasoning models. According to Bloomberg Law the company now believes DeepSeek violated its terms of service by using its application programming interface (API) to train its recently announced R1 model.
The news comes not long after White House AI and crypto czar, David Sacks, told Fox News in an interview on Tuesday it was “possible” DeepSeek “stole intellectual property from the United States.”
“There’s substantial evidence that what DeepSeek did here is they distilled the knowledge out of OpenAI’s models,” Sacks told the outlet.
The AI industry has been raving about DeepSeek’s ability to quickly and cost-effectively train AI models in one year with just $5.6 million. There is an underlying possibility the reason for the company’s efficiency is that it has used another company’s model as its baseline.
DeepSeek may have used a process called distillation, which entails two models having a teacher-student dynamic so one can collect information from the other. On one hand, this could explain the company’s inexpensive operating costs and use of less powerful Nvidia H800 chips. DeepSeek may now be on the hook to prove whether it took all unlawful actions when developing its models.
Before this recent development, industry experts previously speculated that DeepSeek likely used reverse engineering to train its models. This process analyzes models to identify their patterns and biases for improving future models. Reverse engineering is a common practice among open-source developers that is considered legal.
Security researchers sanctioned by Microsoft have already pieced together that DeepSeek may have exhumed a considerable amount of code from OpenAI’s API during the fall of 2024. Microsoft supposedly made OpenAI aware of the breach at the time. The R1 model was announced last week, bringing attention to the Chinese AI company, and associated parties.
DeepSeek has also been lauded as an open-source AI application, on which anyone can develop. This is from where much of the excitement surrounding the platform comes — in addition to its comparison to top tools such as ChatGPT and Google Gemini. OpenAI is not an open-source service; however, anyone can sign up to access its API. The company does make clear in its terms of services that other entities cannot use output to train other AI models, TechCrunch noted.
An OpenAI spokesperson told Reuters that regardless of regulations, various international companies trying to copy models from well-known companies in the U.S. is now a common occurrence.
“We engage in counter-measures to protect our IP, including a careful process for which frontier capabilities to include in released models, and believe as we go forward that it is critically important that we are working closely with the U.S. government to best protect the most capable models from efforts by adversaries and competitors to take U.S. technology,” the spokesperson said.
