Anthropic, China + AI policy by Toju Duke, a Top AI Expert, Ex-Google

4 Mar 2026 · 2 min

Watch on YouTube

1:31

In this monologue, the speaker outlines a growing controversy over AI ownership, data mining, and regulation. Anthropic has accused three leading Chinese AI companies—Deepseek, Moonshot, and Minimax—of illegally using Anthropic's Claude model to improve their own models via massive distillation, allegedly generating over 16 million exchanges through about 24 000 fraudulent accounts. The speaker explains distillation as a common technique where a smaller model is trained on the outputs of a larger one, noting its legitimate uses but highlighting the alleged misuse. He references the 2025 launch of Deepseek as a competitive challenge to U.S. AI dominance and recalls that OpenAI also accused Deepseek of building on GPT‑4 and O1. The broader point is that the AI ecosystem relies heavily on scraping publicly available data, raising ethical concerns about the fair use and compensation of the human subjects who own that data. The speaker urges coordinated international policy efforts to ensure fair compensation and responsible data practices.

Chapters

The controversy on AI ownership, data mining, AI laws and regulations are compounding. Anthropic recently accused some of the leading AI companies in China, Deepseek, Moonshot, and Minimax, of illegally using Claude to improve their models. According to Anthropic, the Chinese labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts using a common machine learning technique called distillation. Distillation involves training a less capable, a large model on the outputs of a larger and more capable one, a very common method in AI development where AI labs and companies use distillation on their own models to create smaller and cheaper models and sometimes more capable models. This takes me back to the beginning of 2025 when the Deepseek model was first released, posing as a good competitor to the US dominance on AI, which I discussed on China's Global network, CGTN, at the time of the release. OpenAI also accused Deepseek during this time of building their models from GPT-4 and O1. This is a bit funny and at the same time interesting and controversial, as the entire AI ecosystem is built from scraping people's data off the internet. And I think it's about time policymakers make a concerted effort across countries to ensure that the fair use of people's data for AI development includes and compensates the human subjects who own the data.