US AI Giants Accuse Chinese Firms of Industrial-Scale Model Distillation
Anthropic and OpenAI report systematic efforts by Chinese labs to bypass restrictions and extract reasoning capabilities from frontier models.
Frontier AI labs in the United States are sounding the alarm over systematic efforts by Chinese competitors to harvest proprietary intelligence through a process known as model distillation. Anthropic and OpenAI have identified large-scale campaigns designed to extract high-value capabilities from their most advanced models, escalating a geopolitical struggle over AI supremacy.
In February 2026, Anthropic revealed that Chinese AI firms DeepSeek, Moonshot, and MiniMax conducted industrial-scale distillation attacks against Claude. These entities allegedly utilized fraudulent accounts and sophisticated routing techniques to bypass regional restrictions and terms-of-service prohibitions. OpenAI has similarly informed U.S. lawmakers that it detected unauthorized attempts by Chinese entities, specifically DeepSeek, to pull outputs from its frontier models for the purpose of distillation.
The Mechanics of Distillation
Model distillation is a machine-learning technique where a larger "teacher" model's outputs are used to train a smaller "student" model. This process allows developers to create AI that is cheaper and easier to deploy while retaining much of the original model's performance.
According to reports, the primary target of these campaigns is not the final answer provided by the AI, but the "reasoning traces" or chain-of-thought. These traces transfer the actual problem-solving methodologies of the teacher model to the student. Florian Tramèr, an assistant professor at ETH Zurich, explains the value of this data by noting that a student has a much harder time learning to solve problems from a book of final solutions than from detailed solutions that describe every step taken.
A Geopolitical End-Run
This technical conflict is deeply rooted in U.S. trade policy. Strict export controls on advanced GPUs have made it increasingly difficult for Chinese firms to train frontier-scale models from scratch due to a lack of compute power. Distillation provides a strategic workaround, allowing these firms to create competent "knock-offs" of American technology for a fraction of the compute cost by simply querying existing APIs.
U.S. officials view these practices as an end-run around chip export controls, transforming a terms-of-service violation into a matter of national security. The ability to synthesize the reasoning capabilities of a model like Claude or GPT-4 without the requisite hardware undermines the intended effect of the hardware sanctions.
The Future of API Access
The fallout from these revelations is likely to trigger a bifurcation of the AI ecosystem, pushing the industry toward increasingly guarded, opaque closed models versus open-weight alternatives.
For the broader developer community, this shift suggests a future of significantly tighter restrictions. Industry analysts expect stricter account verification, more aggressive rate limits, and the suppression of reasoning traces in API responses. While these measures aim to protect intellectual property, they will likely increase the cost and technical difficulty for legitimate developers attempting to build specialized AI assistants.