Key Highlights
- Claims that Chinese AI firms are copying American technology through a process called distillation have become a central issue in talks between President Trump and Chinese President Xi Jinping.
- Distillation involves Training a new AI model using outputs generated by an existing, more advanced system, a technique with legitimate origins in AI research.
- Legal experts say the practice occupies uncertain legal territory, with U.S. courts yet to determine whether it violates Trade secrets law.
- Even where distillation occurs, experts say it cannot fully explain how competitive Chinese AI models have become, since distillation alone does not provide access to a system's underlying code.
A Technical Dispute Turns Geopolitical
For nearly two years, American AI companies including Anthropic, OpenAI and Google have accused Chinese firms of unfairly copying their technology using a process known as distillation. As Chinese President Xi Jinping visited President Trump this week in Washington, the dispute has become a flashpoint in discussions between the two governments, arising amid broader concern over the rapid advancement of AI and calls from officials, executives and members of the public for the technology's development to slow down.
American companies such as Anthropic and OpenAI currently lead the global AI race, but Chinese start-ups including Z.ai and Moonshot are not far behind, having narrowed the gap considerably over the past year. Silicon Valley's repeated claim that Chinese firms are using distillation to close that gap has become one of the more contentious issues on the table as the two leaders meet.
Some experts argue the claims are overstated. A head of model training at Baseten, a U.S. company that sells access to Chinese AI technologies, said the narrative attributing most Chinese capability gains directly to distillation from American models is not as accurate as commonly portrayed.
What Distillation Actually Involves
Distillation was first developed in the early 2010s as a technique for building more efficient AI systems. Researchers would collect data from an established, resource-intensive system and use that data to train a new system capable of running on considerably less expensive hardware. Geoffrey Hinton, a former Google researcher who helped develop the technique, has described the relationship as one model acting as a teacher and the other as a student.
Companies continue to use distillation for this original purpose. More recently, however, the technique has also become a method for companies to extract data from technologies built by competitors. An outside company seeking to distill a rival's system will typically create numerous accounts on that competitor's service, such as Anthropic's Claude or OpenAI's GPT, then assign those accounts to specific tasks like computer coding or mathematics. The distilling company can then use the resulting outputs to help train its own new system.
A Legally Uncertain Practice
Some legal scholars argue that distillation could violate the Defend Trade Secrets Act, a law permitting companies to sue over the theft of trade secrets, though U.S. courts have not yet resolved this question. Copyright law does not necessarily apply in most cases, since distillation generally involves copying a system's general behavior rather than reproducing its underlying text verbatim.
Critics of Anthropic and OpenAI note that both companies have faced similar accusations regarding their own training practices. Anthropic is currently facing multiple lawsuits alleging it illegally used copyrighted internet data to train its systems, and the company agreed last year to pay $1.5 billion to a group of authors and publishers after a judge ruled it had illegally downloaded and stored millions of copyrighted books, marking the largest settlement in the history of U.S. copyright cases. OpenAI and Microsoft face a similar lawsuit brought by The New York Times in 2023, alleging the companies used millions of Times articles to train Chatbots that now compete with the publication as an information source; both companies have denied the claims.
Evidence of Chinese Distillation Remains Unclear
Chinese firms have most likely distilled proprietary American systems to some degree, though the precise extent remains unclear. After the Chinese start-up DeepSeek released a notably powerful and efficient AI system last year, surprising American researchers and investors, OpenAI accused the company of distilling its technology. As Chinese systems have continued improving over the past nine months, Anthropic made similar accusations, sending a letter in June to Senators Tim Scott and Elizabeth Warren accusing Chinese technology giant Alibaba of distilling its models. The Chinese companies named in these accusations have not responded publicly.
Distillation Alone Does Not Explain China's Progress
Experts say distillation cannot fully account for how competitive Chinese AI models have become. When a company distills a proprietary system, it gains access only to the outputs that system generates, not to its underlying computer code, unlike what would be available with an open-source technology. Using outputs from a system such as Anthropic's Claude can accelerate a Chinese company's development process, but those outputs represent only part of what is needed to build a competitive model.
A co-founder of alphaXiv, a Silicon Valley start-up that tracks AI research, said companies distilling externally do not gain access to everything they would have if building a model entirely in-house. Before a Chinese company can effectively use distilled outputs, it must first construct a system that is already substantially capable in its own right, since distillation refines rather than originates a model's core capabilities. That refinement process still requires considerable additional Investment in computing power, talent and engineering work.
Limited Options for Stopping the Practice
Anthropic and OpenAI have attempted to curb distillation, primarily by shutting down accounts suspected of being used for the practice. However, new accounts typically emerge to replace those that are closed, and aggressive account termination risks inadvertently barring legitimate users as well. A researcher at alphaXiv said stopping distillation entirely is essentially impossible given the current structure of how these AI services operate.
Outlook
Whether the distillation dispute produces any concrete agreement between Washington and Beijing will likely depend on how the broader AI safety and trade discussions between Trump and Xi develop in the coming days. Given the unresolved legal status of distillation under U.S. law and the practical difficulty of preventing it, experts suggest the issue is likely to remain a recurring point of friction between the two countries' AI industries regardless of what emerges from this particular summit.






Please wait processing your request...