Artificial intelligence companies in the United States are intensifying their efforts to protect their most advanced models from a technique known as AI distillation, which they say is allowing Chinese competitors to accelerate their technological progress. Anthropic recently urged U.S. lawmakers to investigate the issue, alleging that Alibaba used tens of thousands of unauthorized accounts to access its Claude chatbot and collect data for training competing AI systems. The company described these activities as coordinated, large-scale efforts to extract the capabilities of leading American AI models and repackage them into rival products, tells The New York Times (full article available to subscribers).
Distillation itself is not a new concept. Developed by Google researchers more than a decade ago, the technique allows a smaller AI model to learn from the outputs of a larger, more capable model. Much like a teacher guiding a student, the larger model provides examples that help the smaller model reproduce similar behavior while requiring less computing power. The approach has long been used to improve efficiency and reduce deployment costs, particularly when developers are working with their own AI systems or open-source models.
The controversy arises when companies apply distillation to proprietary AI models without permission. Anthropic and OpenAI prohibit this practice through their terms of service, arguing that it enables competitors to benefit from years of research and billions of dollars in development costs. However, the practice appears to be widespread. During court testimony earlier this year, Elon Musk acknowledged that AI companies generally distill technologies developed by other AI firms, highlighting the industry’s complex competitive landscape.
Legal questions surrounding distillation remain unresolved. Some experts believe unauthorized distillation could violate U.S. trade secret laws, but courts have yet to establish clear legal precedents. Copyright law also provides limited guidance because distillation copies a model’s behavior rather than reproducing its text verbatim.
The debate has intensified as Chinese companies such as DeepSeek and Z.ai rapidly close the performance gap with leading U.S. models. While American companies argue that unauthorized distillation has contributed to China’s progress, many researchers believe it is only one factor. Building frontier AI systems still requires extensive research, engineering expertise, specialized computing hardware, and large-scale training infrastructure. As AI evolves toward autonomous agents capable of performing complex tasks, some experts predict that distillation will play a smaller role, making innovation and original research more important than simply replicating existing models.