Chinese firms expand use of U.S. AI model outputs to train domestic systems
Just the facts

Chinese firms expand use of U.S. AI model outputs to train domestic systems

Summary

Chinese researchers and companies are increasingly employing a technique called distillation, using outputs from leading U.S. AI models to develop smaller, cheaper domestic models as U.S. chip restrictions tighten.

Chinese entities are scaling up the practice of "distillation," where outputs from leading American artificial-intelligence models are used to train smaller domestic systems, according to a review of more than 300 academic papers, patents, corporate disclosures, recruitment notices and government documents. The approach, which has moved from academic experiments to mainstream industrial use, allows Chinese firms to create training data by having large U.S. models answer questions, write code or solve problems, then automating the collection and preparation of that material for "student" models that require less computing power.

The technique is not illegal and is employed worldwide, but it can reduce reliance on foreign technology by preserving selected capabilities in cheaper models. Beijing promotes distillation as a way to make AI more affordable and easier to deploy across its manufacturing sector.

U.S. officials have warned that large-scale distillation could amount to intellectual-property theft. In July, Commerce Department officials raised concerns with Chinese regulators, and in September publicly accused Chinese companies of aggressive industrial-scale distillation. Chinese officials described the method as a standard AI-development practice and said they would act if it crossed into IP theft, though no specific commitments were made.

Examples cited include Shanghai Jiao Tong University researchers who selected about 5,000 answers from OpenAI's o1 model, translated them with GPT-4o mini and created 10,750 bilingual examples to further train Alibaba's Qwen2.5-72B model. Alibaba later reported using GPT-4 and GPT-4o to generate multilingual training material for smaller Qwen models. The papers note that over-reliance on distillation could hinder original innovation.

U.S. AI providers such as OpenAI and Anthropic restrict the use of their outputs for competing models, but Chinese developers argue that distillation is only one factor among advances in engineering, algorithms, data and model design. A national AI-plus-manufacturing policy package released in January by eight Chinese agencies encourages model training, fine-tuning and distillation for industrial applications, reflecting the goal of deploying smaller, cost-effective AI across the economy.

Chinese AI developers have recently sought guidance from industry regulators on the line between ordinary use of foreign models and organized collection of outputs for training, highlighting the growing geopolitical sensitivity of the practice.

Source

Reuters

Keep reading — for free

Create a free account to follow the news. No card required.