Researchers Pull Reasoning Traces From Claude And GPT. The Findings Are Inconvenient. Distillation Is The Real Story.

Researchers found a method to extract hidden reasoning traces from Claude, GPT, and Gemini. The results suggest some Chinese AI models may have been trained on outputs from leading US models. Mark Zuckerberg defended the practice in a blog post, calling distillation "an important principle of how the open source ecosystem works" and warning that restricting it would disadvantage the US. Kyle Miller, a researcher at the Center for Security and Emerging Technologies, says it remains unclear how much distillation actually helps China. OpenAI told US lawmakers in February that DeepSeek appeared to have copied one of its models.

This illustrates the concept of model distillation, where a smaller or competing model learns by absorbing the outputs of a more capable one. The mechanism is called knowledge transfer. It is how students learn from teachers. The geopolitical anxiety is almost amusing. If you cannot prevent information from flowing, you can only decide whether to restrict the source or accept the diffusion. The same principle applies to your own AI usage. You learn faster by studying what better models produce.

The research targets Claude, GPT, and Gemini as source models. Mark Zuckerberg publicly defended distillation in a blog post. Kyle Miller of the Center for Security and Emerging Technologies expressed skepticism about the scale of benefit to China. OpenAI raised concerns about DeepSeek to US lawmakers in February.