Microsoft, a major investor in OpenAI, is investigating whether Chinese AI firm DeepSeek improperly used OpenAI’s data to train its new R1 reasoning model. The investigation centers around potential violations of OpenAI’s API terms of service, specifically regarding the use of API output for training competing AI models.
This inquiry follows recent allegations from White House AI and crypto advisor, David Sacks, who suggested DeepSeek may have “stolen intellectual property” from the United States. Sacks asserted that there’s “substantial evidence” indicating DeepSeek “distilled the knowledge” from OpenAI’s models.
DeepSeek has garnered significant attention for its rapid and cost-effective AI model training, achieving impressive results with a reported investment of just $5.6 million over one year. This efficiency has raised questions about the company’s methods, with speculation arising that DeepSeek might have leveraged another company’s model as a foundation.
One possible technique under scrutiny is “distillation,” a process involving a “teacher-student” dynamic between two models, allowing one to learn from the other. This could explain DeepSeek’s lower operating costs and use of less powerful Nvidia H800 chips. The company now faces the challenge of demonstrating the legality of its development practices.
DeepSeek AI running on an iPhone.
Prior to these developments, industry experts speculated that DeepSeek employed reverse engineering to train its models. This involves analyzing existing models to understand their patterns and biases, which is a common and legal practice among open-source developers.
Microsoft’s security researchers have reportedly found evidence suggesting DeepSeek may have extracted a significant amount of code from OpenAI’s API during the fall of 2024. OpenAI was allegedly informed of this potential breach at the time. The subsequent announcement of DeepSeek’s R1 model last week brought the company and its practices into the spotlight.
DeepSeek’s open-source nature, allowing anyone to develop on its platform, has contributed to its growing popularity, along with comparisons to leading AI tools like ChatGPT and Google Gemini. While OpenAI is not open-source, its API is accessible to registered users. However, OpenAI’s terms of service explicitly prohibit using its output to train other AI models, as highlighted by TechCrunch.
An OpenAI spokesperson told Reuters that attempts by international companies to copy models from established U.S. firms are becoming increasingly common, regardless of regulations. The spokesperson emphasized the importance of collaborating with the U.S. government to protect advanced AI models from unauthorized use by competitors and adversaries. They also highlighted OpenAI’s countermeasures, including careful selection of capabilities included in released models.
In conclusion, the investigation into DeepSeek’s training methods raises important questions about intellectual property in the rapidly evolving AI landscape. The outcome of this investigation could have significant implications for the future development and regulation of AI technologies.
