The violent shock caused by Chinese technology company DeepSeek in the US artificial intelligence industry has lasted for several days, and some different voices are constantly coming out. The Financial Times of London reported on Wednesday (January 29) that OpenAI, the developer of the US artificial intelligence chatbot ChatGPT, pointed out that it has found evidence that the Chinese startup DeepSeek used OpenAI’s proprietary model to train its own open source chatbot. This provides new evidence for those who suspect that the Chinese company may have infringed intellectual property rights.
The Financial Times said that the ChatGPT manufacturer headquartered in San Francisco, USA, said it had seen some evidence that DeepSeek’s developers had infringed through the “distillation” technology.
The so-called “distillation” technology refers to a practice in which developers use more powerful model outputs to train small models, thereby achieving higher performance at a lower cost.
The Financial Times said that this is a common practice in the industry, but the problem is that the intelligent chatbot developed by DeepSeek and OpenAI’s ChatGPT are similar competing products. Its practice is suspected of infringing the product terms of use stipulated by OpenAI and may constitute infringement.
The report quoted a person close to OpenAI as saying, “The problem is that you (take it off the platform) to create your own model for your own purposes.”
Tesla CEO Elon Musk also reportedly pointed out recently that he did not believe DeepSeek’s AI research and development capabilities and suspected that the Chinese company had many more Nvidia graphics cards than it had announced. The
technology information website “TheVerge.com” published an article on Wednesday saying that OpenAI and US technology giant Microsoft are investigating whether the Chinese competitor used OpenAI’s API (application programming interface) to integrate OpenAI’s artificial intelligence model into DeepSeek’s own model. Sources said that Microsoft security researchers discovered that a large amount of data was leaked through the OpenAI developer account at the end of 2024, and Microsoft believed that this account was associated with DeepSeek. The Financial
Times said that OpenAI declined to make further comments or provide detailed evidence. The company’s product terms of use clearly stipulate that users cannot “copy” any of its services or “use its output to develop models that compete with OpenAI.”
The R1 inference model released by DeepSeek surprised the market, investors and technology companies in Silicon Valley, USA. This model is low-cost, highly ranked, and comparable to the leading models in the United States.
DeepSeek’s model was released last weekend, triggering a sharp drop in U.S. technology stocks on Monday. The stock of Nvidia, the U.S. AI chip giant, fell 17%, and its market value evaporated by nearly $600 billion. Investors are worried that the U.S. model of developing AI through huge investments may have a big bubble, which will cause the market to re-evaluate their value. However, on Tuesday, the market’s anxiety was calmed down, with the Nasdaq rebounding more than 400 points and Nvidia’s stock rebounding by nine percentage points.
DeepSeek shock abates, global markets stabilize, Nasdaq rebounds strongly
According to Bloomberg, a U.S. financial media, a person familiar with the matter said that OpenAI and its partner Microsoft investigated an account believed to be DeepSeek using OpenAI’s application programming interface (API) last year and revoked its access because it suspected that it violated the terms of service by using distillation technology.
The Financial Times said that Microsoft declined to comment and OpenAI did not quickly respond to the request for comment. DeepSeek may not have responded to its request for comment due to the Chinese New Year holiday.
David Sacks, an adviser on artificial intelligence and cryptocurrency affairs to U.S. President Trump, said earlier that it was “possible” that intellectual property theft had occurred.
In an interview with Fox News on Tuesday, Sacks said that “there is a technology in artificial intelligence called distillation,” which refers to the ability of a model to draw knowledge from a master model while learning from another model.
The Financial Times quoted Sachs as saying, “There is a lot of evidence that what DeepSeek did was extract knowledge from OpenAI’s model, and I don’t think OpenAI will be happy with that.”
DeepSeek said it used only 2,048 Nvidia H800 graphics cards and spent $5.6 million to train the V3 model with 671 billion parameters, which is just a fraction of the cost of OpenAI and Google to train models of the same size.
Some experts said the model’s generation of responses showed that it had been trained with OpenAI GPT-4 output, which violated its terms of service.
Industry insiders said it is a common practice for artificial intelligence labs in China and the United States to use the output of companies such as OpenAI, which have paid people to teach their models to make more human-sounding responses.
Insiders said that this practice is both expensive and labor-intensive, and small companies usually adopt a “free ride” approach to do this work.
OpenAI said in its latest statement, “We know that (Chinese) companies and other companies are constantly trying to extract (what they need) from the models of leading American artificial intelligence companies.”
This practice highlights the difficulties faced by companies keen to protect their technological advantages, the Financial Times said.