AI distillation lets rivals copy frontier models at lower cost

AI distillation trains one model on the outputs of another, allowing developers to replicate capabilities without building large training datasets from scratch. U.S. companies including Anthropic, OpenAI, and Google have spent billions on data, talent, and computing power to create frontier models. Rivals can now use distillation to produce similar systems at lower cost.

Industry complaints and regulatory calls Anthropic accused Alibaba of creating tens of thousands of fake Claude accounts to harvest model outputs. Sarah Heck, Anthropic's head of policy, wrote that the practice turns U.S. research investments into a subsidy for competitors. OpenAI stated that blending outputs from multiple U.S. models could create systems that surpass any single source model.

Earlier development and current workarounds The method originated in 2015 when Google researchers trained smaller models on their own larger systems. After 2022, companies began applying it to competitors' outputs. Chinese developers facing chip limits have built proxy services that bypass account restrictions.

These services collect prompts and responses that can later be used for training. Anthropic has blocked users in China, required overseas phone numbers, and asked some users for government ID verification. Researchers say the restrictions have increased demand for cheaper distilled open-source alternatives.