Microsoft Sets Its Legal Focus on DeepSeek

Microsoft, a key investor in OpenAI, is currently investigating concerns regarding the Chinese company DeepSeek. The investigation revolves around allegations that DeepSeek may have employed questionable methods to train its popular new reasoning model, known as R1. According to Bloomberg Law, there are suspicions that DeepSeek violated Microsoft’s terms of service by utilizing its application programming interface (API) for training purposes.

This situation gained additional attention after David Sacks, the White House AI and crypto czar, suggested in a Fox News interview that it is “possible” that DeepSeek “stole intellectual property from the United States.” Sacks further emphasized, “There’s substantial evidence that what DeepSeek did here is they distilled the knowledge out of OpenAI’s models.”

The AI community has applauded DeepSeek for its capability to train AI models swiftly and economically, costing approximately $5.6 million over the course of a year. However, the company’s efficiency raises questions about whether it may have based its models on another company’s existing work.

DeepSeek is suspected of using a technique known as distillation, where a new model learns from an existing one, adopting its knowledge in a teacher-student relationship. This approach could be a reason behind DeepSeek's lower operational costs and its reliance on less powerful Nvidia H800 chips. As a result, DeepSeek might now need to demonstrate that its model development processes adhered to legal standards.

Prior to these recent allegations, industry experts had speculated that DeepSeek likely employed reverse engineering techniques to enhance its models. This method involves scrutinizing existing models to understand their patterns and biases, which helps in developing improved future models. Reverse engineering is commonly used among open-source developers and is generally considered a legal practice.

Security researchers working with Microsoft reportedly found that DeepSeek may have extracted a substantial amount of code from OpenAI’s API in the fall of 2024. Microsoft is believed to have notified OpenAI regarding this breach at the time. The announcement of the R1 model last week has drawn more attention to DeepSeek and its operations.

DeepSeek is marketed as an open-source AI platform, inviting development by anyone, which contributes to the excitement around it. In contrast, while OpenAI’s services are not open-source, it does allow users to sign up for access to its API. It is important to note that OpenAI clearly states in its terms of service that other companies cannot use the output generated to train their own AI models, as reported by TechCrunch.

An OpenAI spokesperson communicated to Reuters that the illicit replication of models by various international companies is an increasing occurrence. They stated, “We engage in counter-measures to protect our IP, including a careful process for which frontier capabilities to include in released models. As we go forward, we believe it is critically important to work closely with the U.S. government to safeguard our advanced models from adversaries and competition that aim to replicate U.S. technology.”

Microsoft, DeepSeek, API