The AI technique at the centre of a US-China technology dispute, explained

Key points

• Chinese lab Moonshot AI released Kimi K3, an open AI model that users found competitive with top US offerings from Anthropic and OpenAI.
• White House advisor Michael Kratsios posted on 25 June 2025 that Moonshot used Anthropic's Fable model without permission to build K3.
• Anthropic reported in February 2025 that Chinese firms used roughly 24,000 fake accounts to generate 16 million exchanges with its Claude models, apparently to copy their capabilities.
• More than 20 tech companies, including Nvidia, Microsoft and Meta, signed a letter urging policymakers not to restrict open AI models over distillation concerns.
• Both Anthropic and OpenAI ban distillation in their terms of service, but US companies have also faced lawsuits for training their own models on content they did not own.

A Chinese AI lab just embarrassed some of Silicon Valley's biggest names, and the method it may have used to do it is now at the centre of a fight between the tech industry and Washington.

The lab is Moonshot AI. Its new model, Kimi K3, is what researchers call an open-weight model: software you can download, modify and run on your own computers, rather than rent time on through a company's website. When people tested Kimi K3 last week, they found it matched or beat models from Anthropic and OpenAI, the two US companies widely considered to lead the field.

That result alarmed US officials. The question they are asking: how did Moonshot catch up so fast?

What is distillation, and why does it matter?

Distillation is how a smaller, cheaper AI model learns from a bigger, more expensive one. Think of it like a student copying a classmate's homework instead of doing the reading themselves. The smaller model feeds questions into a frontier model, the most advanced version of a given AI system, then trains itself on the answers it gets back.

The result can be a capable model built for a fraction of the original cost.

Google AI head Jeff Dean described the technique approvingly on a podcast in February, noting it was central to how Google improved its own products. Nvidia used distillation to build its Llama Nemotron series of models. The practice is, as one industry researcher put it, "legitimate and very valuable".

The controversy is about who is doing it, and whether they had permission.

Did China steal American AI?

White House advisor Michael Kratsios said plainly on social media that Moonshot distilled Anthropic's Fable model to build K3, adding that the lab built a platform specifically designed to switch between access methods to avoid being caught.

Anthropic had already raised this alarm. In February 2025, the company said Chinese firms DeepSeek, Moonshot and MiniMax used around 24,000 fake accounts to run 16 million conversations with its Claude AI, apparently harvesting those responses to train rival models. Anthropic, currently valued at close to one trillion dollars, called stopping this practice a matter of national security.

Both Anthropic and OpenAI now explicitly ban distillation in their terms of service.

Company	Position on distillation	Key action taken
Anthropic	Bans it as IP theft	Terms of service prohibition, public warning
OpenAI	Bans it as IP theft	Terms of service prohibition
Nvidia	Uses it in own models	Signed letter defending the practice
Microsoft, Meta	Defend it as standard	Signed letter urging no restrictions
Moonshot AI	Disputed	Released open-weight Kimi K3

What did the tech industry say?

More than 20 companies pushed back hard. Nvidia, Microsoft, Meta and Palantir joined a letter asking policymakers not to restrict open-weight AI models, arguing distillation is a standard engineering tool and that banning it would hurt American companies more than Chinese ones.

Box chief executive Aaron Levie, who signed the letter, said US companies need access to the best technology wherever it comes from.

There is also an uncomfortable wrinkle for Anthropic and OpenAI. Both face ongoing lawsuits from authors and other creators who say their work was used without permission to train the very frontier models now being protected. A copyright attorney representing book authors told CNBC Tech the US administration has focused on protecting AI companies' intellectual property while staying quiet about creators whose work those same companies used first.

What does this mean for ordinary people?

For most users, nothing changes immediately. Kimi K3 is free to download and use. The broader fight is about whether cheap, capable AI models, many originating from China, will be restricted or taxed in the US market.

If distillation rules tighten, the cost of building AI products could rise, potentially pushing prices up for businesses that rely on AI tools. If they do not tighten, expect more capable Chinese models at lower prices, raising separate questions about data privacy and security.

Pukar Hamal, who runs AI security firm SecurityPal, said he would use Kimi K3 if he could verify no hidden backdoors existed in the code. That is sensible advice for any business considering open-weight models from any source: check what the software is actually doing before you run it on sensitive systems.

Common questions

Is distillation always illegal?

No. Using distillation on publicly available models, or with permission, is legal and common. The dispute centres on using a company's paid service to harvest data for a competing product, which Anthropic and OpenAI explicitly ban in their terms.

Should I be concerned about using Chinese AI models?

Not necessarily, but caution is reasonable. Independent security testing before deployment is good practice with any third-party AI software, regardless of origin.

Could this lead to new laws restricting AI?

It is possible. US policymakers are actively debating the question, and the outcome will affect which AI tools businesses and consumers can access and at what cost.