AI Distillation Controversy: Did China Really 'Steal' American AI?

In February 2026, a series of explosive revelations rocked the AI industry. OpenAI formally accused DeepSeek of 'systematic free-riding' before the U.S. House Select Committee on China, while Anthropic published evidence that three Chinese AI firms, DeepSeek, Moonshot AI, and MiniMax, had conducted large-scale distillation of its Claude model. Behind the technical term 'knowledge distillation' lay a shocking reality: over 20,000 fraudulent accounts, more than 16 million data exchanges, and the illegal use of export-banned chips.

1. Knowledge Distillation: How AI Copies AI

Knowledge distillation is a technique that trains smaller AI models using the outputs of larger ones. While it is a legitimate machine learning method, problems arise when commercial models are distilled at scale without authorization. The method involves sending tens of thousands of prompts to a competitor's API and collecting the responses as training data for one's own model.

The scale of this incident far exceeds casual scraping. According to Anthropic, approximately 24,000 fraudulent accounts were deployed, with over 16 million query-response exchanges recorded. DeepSeek accounted for roughly 150,000 exchanges, Moonshot AI for 3.4 million, and MiniMax for a staggering 13 million.

2. Hydra Clusters: The Infrastructure of Systematic Distillation

The attack methods Anthropic revealed were remarkably sophisticated. A proxy network dubbed 'Hydra Clusters' managed over 20,000 fraudulent accounts simultaneously while evading detection. When one account was blocked, a new one immediately took its place, like the mythological hydra regrowing its severed heads.

OpenAI confirmed similar tactics. DeepSeek employees had been bypassing access restrictions through 'obfuscated third-party routers.' Particularly noteworthy were the prompts DeepSeek sent to Claude, requesting it to 'imagine the hidden internal reasoning behind a completed response and write it out step by step.' This was analyzed as a method to generate massive amounts of Chain-of-Thought training data.

MiniMax's behavior was equally striking. Whenever Anthropic released a new model, MiniMax would redirect half its total traffic to that model within 24 hours, initiating concentrated distillation. The intent to absorb the latest model's capabilities as quickly as possible was unmistakable.

3. Export Ban Violations: The Nvidia Blackwell Chip Controversy

An even more serious issue emerged on top of the distillation controversy. On February 24, Reuters reported that DeepSeek had trained its AI models using Nvidia's latest Blackwell chips, which the U.S. government has banned from export to China. How DeepSeek obtained these chips remains unclear, but the revelation raises fundamental questions about the effectiveness of export controls.

The chip export policy itself was already controversial. In December 2025, the Trump administration allowed the sale of Nvidia H200 chips to China, drawing sharp criticism from congressional hawks. The episode laid bare America's dilemma of balancing technology containment with market interests.

4. U.S. Government and Industry Response: Legislation and Formal Accusations

The American response has been escalating rapidly. NASA blocked DeepSeek, New York State banned it from government devices, and the House introduced the 'No DeepSeek on Government Devices Act.'

OpenAI submitted a memo to the House Select Committee on China in February 2026, formally accusing DeepSeek of 'systematic free-riding.' White House AI advisor David Sacks stated there was 'substantial evidence that DeepSeek distilled knowledge from OpenAI,' while Select Committee Chair Moolenaar characterized it as 'the CCP playbook: steal, copy, kill.'

Anthropic raised security-level concerns as well. Illegally distilled models are unlikely to preserve the safety guardrails of the original, meaning harmful content filtering and bias prevention features may be removed or weakened during the distillation process.

5. China's Position and the Legal Irony

DeepSeek declined to comment on the controversy. The Chinese embassy issued only a statement opposing 'ideological line-drawing,' without offering specific rebuttals.

The legal irony of this dispute is fascinating. OpenAI itself has been sued by the New York Times for copyright infringement and is defending with a 'Fair Use' argument. If training models on web-collected data qualifies as fair use, why is using another AI's output as training data different? CrowdStrike co-founder Dmitri Alperovitch said the distillation has 'now been confirmed as fact,' but the legal boundaries remain blurry.

Closing: Walking the Tightrope Between Tech Supremacy and Ethics

The AI distillation controversy has become the front line of the U.S.-China technology supremacy competition, transcending a mere technical dispute. One side invests hundreds of billions of dollars developing foundation models while the other effectively free-rides on those achievements. Yet in an AI industry where the boundaries of 'fair use' remain ambiguous, distinguishing true innovators from imitators is far from simple.

What is certain is that this controversy is changing the rules of the game. Detection system upgrades, export control overhauls, and new intellectual property frameworks are all being advanced simultaneously. How the industry strikes a balance between tech supremacy and ethics will determine the trajectory of AI development for years to come.