In the past two years, companies such as Anthropic, OpenAI and Google have frequently accused Chinese companies of unfairly misappropriating their artificial intelligence technology. As Chinese President Xi Jinping visits Washington this week to meet with President Trump, the allegation has become a focal point in talks between the two sides.
The leaders of the two countries are expected to discuss the rapid rise of the technology amid growing concerns about the potential risks of AI, while many U.S. officials, technology executives and ordinary people have called for a slower pace of development. Although American companies such as Anthropic and OpenAI are leading the AI race, Chinese start-ups represented by Zhipu and Dark Side of the Moon are also catching up.
One of the most contentious issues in the talks was the repeated accusation from Silicon Valley that Chinese companies are appropriating America's leading AI technology using a technique called "distillation."
Many experts have pointed out that the above claims made by companies such as Anthropic and OpenAI are exaggerated. Charles O'Neill, head of model training at Baseten, an American company that sells Chinese technology interface services, said: "The statement that 'all the capabilities of Chinese technology come from the Anthropic model' is far from true as people say."
This article aims to explain how a technological process unknown to the general public can become a geopolitical issue.
What is distillation?
Distillation technology was first developed in the early 2010s as a means for AI researchers to build more efficient technologies. The technology allows researchers to extract data from mature systems and use it to build new systems that can run on lower-cost hardware.
Jeffrey Hinton, a former Google researcher who helped develop the technology, previously explained in an interview with the New York Times: "You can think of one of the models as the teacher and the other as the student."
Today, various companies still use this distillation method. But in recent years, the technology has also evolved into a way for companies to collect data from technology built by others.
In short, outside companies intent on distilling rival systems would register large numbers of accounts on each other's services (such as Anthropic's Claude or OpenAI's GPT) and then earmark those systems for specific tasks, such as computer programming or mathematical operations.
The distiller can then use the information generated by these mature AI systems to assist in training its own new system.
Is distillation illegal?
Some legal scholars have pointed out that distillation may violate the Trade Secrets Protection Act, which allows companies to sue for trade secret theft. However, U.S. courts have yet to make a clear ruling on this issue.
Copyright law does not necessarily apply in such cases, since Distillation aims to copy the overall behavioral logic of the system rather than copying its text word for word.
Critics of Anthropic and OpenAI point to similar behavior by the companies themselves.
Anthropic is accused of illegally using copyrighted internet data to train its systems and is currently facing multiple lawsuits. Last year, the company agreed to pay $1.5 billion to a group of authors and publishers in a landmark legal settlement after a judge found it illegally downloaded and stored millions of copyrighted books. The settlement amount is the largest in the history of U.S. copyright litigation.
OpenAI and other companies have faced similar lawsuits, including one filed by the New York Times in 2023. The lawsuit accuses OpenAI and its partner Microsoft of using millions of articles published by The New York Times to train chatbots that are now competing with news outlets as sources of information. OpenAI and Microsoft deny the accusations.
Are Chinese companies really appropriating American systems?
Chinese companies most likely distilled U.S. proprietary systems, but the details of their operations are unknown.
After Chinese startup DeepSeek shocked U.S. researchers and investors last year by releasing a system that was both efficient and powerful, OpenAI publicly accused the company of distilling its technology.
Anthropic has made similar accusations as the performance of Chinese systems has continued to improve over the past nine months. In June, Anthropic sent a letter to Senators Tim Scott and Elizabeth Warren accusing Chinese tech giant Alibaba of distilling its technology.
In this regard, the relevant Chinese companies have not yet made a public response.
Is distillation alone enough to explain how China’s big models remain competitive?
The answer is no. When companies distill proprietary systems, what they can obtain is limited to the data generated by the system (such as words, characters, and other output content), and they cannot access the underlying source code of the system like they do with open source technologies.
Using the output of proprietary systems such as Anthropic's Claude Fable, Chinese companies can indeed accelerate the development process of their own technology, but these outputs are only part of the entire technology picture.
Rihan Ahmed, co-founder of alphaXiv, a Silicon Valley startup that tracks the latest AI research, said: "The underlying elements that you can use when distilling your own model internally are simply not available through this external method."
Before using these output contents, Chinese companies must first independently build a system with considerable strength in itself. Subsequently, although companies can polish the model in key dimensions through distillation, this still requires investing more energy, funds, computing power, and talents.
Can Anthropic and OpenAI stop other companies from implementing distillation?
