Artificial intelligence (AI) evolves incredibly quickly. Our understanding about AI has to evolve almost as fast. One of the concepts we’ll have to get up to speed on is distillation.

Simply put, distillation is a way of quickly improving AI models by querying more advanced models, recording the results, and using those advanced AI outputs as a roadmap that can be copied.

Distillation is in the news because Anthropic alleges that China’s Moonshot AI engaged in intellectual property (IP) theft, using nearly 25,000 fake accounts to harvest millions of responses from Claude to train its own Kimi K3 model. Beyond this specific incident, policymakers worry that distillation might be helping China close the gap with U.S. frontier AI (the cutting-edge AI models) faster than it otherwise would.

Distillation reflects a shift in the economics of AI. Rather than investing billions of dollars to train frontier models, developers can build highly capable systems by using the outputs of existing ones. That means lower development costs and more AI competition. Some businesses will probably find that cheaper models are good enough for most of their needs and be unwilling to pay a higher cost for frontier models. Not to mention, most Chinese models are open source – meaning anyone can download, inspect, modify, and run them in their systems – putting greater competitive pressure on private U.S. AI models.

Because distillation helps Chinese AI labs continue to release very capable models at lower prices, policymakers are beginning to worry about what it might mean for continued U.S. leadership in AI, as well as national security. Treasury Secretary Bessent recently suggested the United States could sanction foreign companies that have unlawfully distilled U.S. AI models. Meanwhile, policymakers debate whether current export controls and existing legal frameworks are sufficient to protect U.S. AI innovation from these techniques.

As the broader conversation on AI leadership, IP, and national security unfolds, policymakers should be careful not to treat distillation itself as the problem. Distillation is widely used across the AI industry to improve model performance, so restricting the practice broadly could unintentionally slow innovation and hurt American developers. What matters is how distillation gets used: drawing a line between legitimate distillation and companies obtaining proprietary model capabilities without authorization. After all, distillation isn’t going away – it is becoming a core part of building AI systems – and U.S. leadership won’t be preserved by banning techniques that make AI better.