What happens when an AI learns by copying another AI

On July 16, a Chinese company called Moonshot AI released an AI model called Kimi K3 — the largest open-weight model ever released, meaning anyone can download and run it. It performed well in tests, winning code-writing comparisons against other AI tools. Sign-ups closed within two days because demand was so high.

Six days later, the White House accused Moonshot of building K3 by secretly copying a rival. The technique is called distillation: you train a weaker AI by feeding it the outputs of a stronger one — like a student who learns by copying every answer from the smartest person in class. The White House says Moonshot ran Anthropic's top model, Fable, millions of times and used those answers to teach K3 how to behave. Anthropic traced 3.4 million Claude conversations back to Moonshot before K3 ever launched. Technical tests show K3 sometimes calls itself Claude — a sign it absorbed Fable's identity along with its skills.

The accusation carries real weight. Treasury Secretary Scott Bessent said 'open source is not open season on American IP' and warned that sanctions and export blacklists are on the table. Anthropic's head of policy called it 'IP theft and industrial espionage.' The US government also claims Moonshot obtained Nvidia's most powerful chips — which export rules ban from reaching China — by routing them through Thailand. Moonshot has not responded publicly. Some researchers doubt the timeline: Fable was only publicly available for 15 days before K3 launched.

Whether or not this specific allegation holds, it points to a real risk every AI company faces. If top models can be copied, competitive edges shrink. For businesses that have invested in a specific AI provider, this raises a question about how durable those advantages really are. Expect AI providers to tighten access controls and detection systems — which affects every business running AI at scale.