AI Distillation Sparks National Security Debate Over China

A machine learning technique that Google's AI chief Jeff Dean casually mentioned on a podcast in February has exploded into a full-blown policy controversy, pitting Silicon Valley against Washington over how to balance innovation with national security.

Distillation — the practice of using outputs from advanced AI models to train smaller, more efficient ones — became a hot-button issue after Chinese startup Moonshot AI released its Kimi K3 model in late July. Users quickly found it competitive with leading American models from Anthropic and OpenAI, despite Moonshot offering it as an open-weight system that anyone can download and modify.

White House advisor Michael Kratsios accused Moonshot of distilling Anthropic's frontier Fable model, calling it intellectual property theft. "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," Kratsios wrote on X, adding that the company built a sophisticated platform to conduct large-scale distillation while evading detection.

Why it matters

The distillation debate forces policymakers to confront a fundamental tension: the same technique that helps American companies build cost-effective AI systems can also help foreign competitors rapidly close capability gaps. With AI infrastructure costs soaring and China aggressively pursuing AI leadership, how the U.S. handles distillation could shape both domestic innovation and international competition for years.

Tech giants push back

More than 20 companies — including Nvidia, Microsoft, Meta, and Palantir — issued a joint letter urging policymakers to avoid "premature restrictions" on open-weight models. The coalition argued that distillation is "a widely used technique for model improvement, evolution, and validation," and warned that restrictions could stifle competition or push innovation overseas.

Box CEO Aaron Levie, a signatory, told CNBC that U.S. companies need access to the best technology regardless of origin. "Generally the arc is going to be that the more innovation that there is, whether that's from the U.S. or China or otherwise, you should expect more AI progress," Levie said.

How distillation works

At its core, distillation involves using a powerful "teacher" model to generate training data for a smaller "student" model. Pukar Hamal, founder of AI security firm SecurityPal, compared it to copying homework: "It's almost like someone went to the lectures, read the textbook, and did all the hard work of doing the homework. Then some other student is like, 'Hey, I didn't do that. Can I just copy your work?'"

Nvidia has openly used distillation for its Llama Nemotron models, documenting the approach in research papers. Shashi Bellamkonda, research director at Info-Tech Research Group, noted that "it is a legitimate and a very valuable technique to train a smaller, cheaper model on outputs of a larger model, and is practiced all the time."

The IP problem cuts both ways

Anthropic reported in February that Chinese companies DeepSeek, Moonshot, and MiniMax used approximately 24,000 fake accounts to generate 16 million exchanges, distilling Claude capabilities on an "industrial scale." Both Anthropic and OpenAI now ban distillation in their terms of service.

But their position is complicated by ongoing copyright litigation. Max Pritt, an attorney representing book authors suing AI firms, pointed out the irony: "The administration, at least publicly, has focused its efforts on the protection of technology companies' intellectual property, while remaining silent in large part about creators and individuals' intellectual property that was used without authorization."

Colin Shea-Blymyer, a research fellow at Georgetown's Center for Security and Emerging Technology, said the government is still formulating its stance on whether Chinese companies using American model outputs constitutes an unfair advantage.

These details were first reported by CNBC.