Policy makers and industry leaders have voiced alarm about the rapid expansion of artificial intelligence, with many U.S. officials, technology executives and ordinary citizens urging a slower pace. American firms such as Anthropic and OpenAI dominate the current race, while Chinese newcomers including Z.ai and Moonshot are quickly narrowing the gap.
Some analysts argue that the narrative of Chinese firms stealing U.S. capabilities is exaggerated. Charles O’Neill, head of model training at Baseten, told reporters that the idea that all Chinese advances stem from Anthropic-style models is not as accurate as it appears. The practice of model distillation dates back to the early 2010s, originally intended to create more efficient systems.
Distillation works by treating one model as a teacher and another as a student, a description Geoffrey Hinton gave to The New York Times. Companies now employ the method to harvest outputs from rival services, creating multiple accounts on platforms such as Anthropic’s Claude or OpenAI’s GPT and assigning them tasks ranging from code generation to mathematical problem solving. The collected responses are then used to train new, independent models.
Legal scholars note that this approach could run afoul of the Defend Trade Secrets Act, which permits lawsuits over the misappropriation of proprietary information, although U.S. courts have yet to issue a definitive ruling. Copyright law is less clear-cut because distillation copies the behavior of a system rather than reproducing its exact text, making traditional infringement arguments harder to apply.
Both Anthropic and OpenAI have faced accusations of similar conduct. Anthropic is defending multiple lawsuits alleging illegal use of copyrighted internet material, and last year it settled for $1.5 billion with authors and publishers after a judge found it had improperly downloaded millions of books. OpenAI and its partner Microsoft are contesting a 2023 lawsuit filed by The Times, which claims the firms used millions of the newspaper’s articles to train competing chatbots.
Chinese firms are suspected of applying distillation to proprietary American models, though the exact scope remains uncertain. OpenAI publicly accused DeepSeek of copying its technology after the Chinese startup released a notably efficient system. In June, Anthropic sent a letter to Senators Tim Scott and Elizabeth Warren alleging that Alibaba had engaged in similar distillation practices.
When a company distills a proprietary system, it only receives the generated text, not the underlying source code, limiting the depth of the copied knowledge. Rehaan Ahmad, co-founder of the Silicon Valley tracker alphaXiv, explained that a Chinese developer must already possess a strong base model before distillation can meaningfully accelerate progress, a process that still demands substantial funding, compute power and expertise.
Providers can block accounts they believe are being used for distillation, but new accounts often appear, and overly aggressive shutdowns risk excluding legitimate users. Lino Le Van, another researcher at alphaXiv, warned that halting the practice is essentially impossible, underscoring the difficulty of enforcing intellectual-property protections in the fast-moving AI landscape.