Anthropic Reveals Chinese AI Companies Used Distilled Claude Data to Boost Local Models

Anthropic reports that Chinese AI companies "distilled" its Claude models to improve their own systems, highlighting global tensions in AI development.

By: AXL Media

Published: Feb 24, 2026, 5:42 AM EST

Source: Information for this report was sourced from NBC News

The Strategy of Model Distillation

Model distillation is a common technique in the AI industry where the outputs of a "teacher" model in this case, Anthropic’s Claude are used to train a "student" model. This allows the student model to mimic the reasoning and accuracy of the more powerful system without requiring the same level of computing power. While this is often used legitimately within companies, Anthropic’s report suggests that Chinese firms accessed Claude via third party interfaces or regional APIs to harvest these high quality responses and accelerate their own development timelines.

Navigating Export Controls and Access

The discovery comes at a time of heightened tension between the US and China over the transfer of advanced technology. While the US has implemented strict hardware restrictions on high end chips, regulating access to cloud based AI software remains a significant challenge. By using proxies or international cloud providers, developers in China can still interact with Western models. This allows them to bypass certain barriers and use the refined linguistic capabilities of Claude to fine tune local models like those from Alibaba or Baidu.

Intellectual Property and Ethical Concerns

Anthropic’s findings have sparked a debate over the protection of AI intellectual property. Unlike traditional code, the "knowledge" stored in an AI model’s weights can be partially extracted through its outputs. Anthropic has stated that it monitors for "automated scraping" and unusual patterns that suggest a model is being used for training purposes rather than human interaction. The company is reportedly strengthening its terms of service and technical safeguards to prevent competitors from "piggybacking" on their expensive research and development.