A once little-known artificial intelligence training method known as distillation has become the center of a growing debate among technology companies and U.S. policymakers, following allegations that Chinese AI firms are using the technique to rapidly narrow the gap with leading American developers.
The discussion gained momentum after Chinese AI company Moonshot AI introduced its Kimi K3 model, which many users have described as performing at a level comparable to the latest commercial offerings from U.S.-based companies including OpenAI and Anthropic. Unlike many proprietary American models, Kimi K3 is released as an open-weight model, allowing developers to download, modify, and deploy it on their own infrastructure.
The model’s release has reignited concerns in Washington over whether AI distillation is enabling Chinese firms to benefit from years of costly research conducted by U.S. companies.
White House science and technology adviser Michael Kratsios alleged that Moonshot AI used Anthropic’s frontier AI model, Fable, during the development of Kimi K3. In a post on X, Kratsios claimed the company created an internal platform capable of conducting large-scale distillation while switching between multiple access methods to avoid detection. Moonshot AI has not publicly responded to the allegation.
What Is AI Distillation?
Distillation is a machine learning technique in which a smaller AI model is trained using the outputs generated by a more advanced model. The approach allows developers to build efficient systems that require fewer computing resources while retaining much of the performance of larger models.
Google’s Chief Scientist Jeff Dean highlighted the importance of the technique earlier this year, explaining that Google adopted distillation to improve the capabilities of smaller AI models without relying exclusively on massive neural networks.
Industry experts say the process has become a standard part of AI development.
“It is a legitimate and valuable technique to train a smaller, more efficient model using the outputs of a larger one,” said Shashi Bellamkonda, Research Director at Info-Tech Research Group, noting that companies including Nvidia have publicly documented the use of distillation in their own AI models.
Growing Divide Over Intellectual Property
While many technology firms view distillation as a normal engineering practice, companies developing frontier AI systems argue that unauthorized use of their models crosses into intellectual property infringement.
Anthropic has accused Chinese companies including DeepSeek, Moonshot AI, and MiniMax of conducting AI distillation at an industrial scale. Earlier this year, the company alleged that approximately 24,000 fake accounts generated millions of interactions with its Claude models to extract training data.
OpenAI and Anthropic have both updated their terms of service to prohibit customers from using their models to train competing AI systems.
AI security expert Pukar Hamal, founder of SecurityPal, compared the practice to copying another student’s completed coursework after someone else had invested the effort to learn the material.
Tech Industry Calls for Balanced Regulation
The controversy has also exposed divisions between AI developers and the broader technology industry.
A coalition of more than two dozen companies, including Nvidia, Microsoft, Meta, Palantir, and Box, recently urged U.S. policymakers to avoid imposing restrictions on open-weight AI models. In a joint letter, the companies argued that distillation remains a widely accepted method for improving AI systems and warned that excessive regulation could reduce competition while pushing innovation overseas.
Box CEO Aaron Levie said access to the world’s best AI technology, regardless of where it is developed, will remain essential for maintaining competitiveness as models become more efficient and affordable.
National Security Concerns Continue
U.S. officials are now weighing how to address the issue as AI becomes increasingly tied to economic competitiveness and national security.
Researchers at Georgetown University’s Center for Security and Emerging Technology say policymakers face a difficult challenge. On one hand, they seek to protect American technological leadership and intellectual property. On the other, many of the same AI techniques under scrutiny are routinely used by U.S. companies.
The debate is further complicated by ongoing copyright lawsuits against several leading AI developers, including OpenAI and Anthropic, over allegations that copyrighted books and other creative works were used without authorization during model training.
Legal experts argue that any discussion about protecting AI companies’ intellectual property must also address the rights of creators whose content contributed to building today’s advanced AI systems.
