U.S. AI Labs Accuse Chinese Firms of Model Theft via Distillation

Leading U.S. artificial intelligence companies are raising alarms that Chinese competitors are systematically copying their technology through a process called distillation, using tens of thousands of unauthorized accounts to harvest proprietary AI capabilities.

In a June 10 letter to Senators Tim Scott and Elizabeth Warren, Anthropic accused Chinese tech giant Alibaba of surreptitiously accessing its AI systems to train competing models. The company characterized the practice as occurring "illicitly, systematically and at industrial scale" across multiple frontier AI labs, according to the letter obtained by The New York Times.

How distillation works

Distillation is a technique where a smaller AI model learns to mimic a larger, more capable system by studying its outputs. Chinese companies have allegedly been creating thousands of unauthorized accounts to query American AI systems, collecting the responses to train their own models without the computational expense or research investment required to build frontier systems from scratch.

Anthropic asked the senators, who lead a committee that recently held hearings on AI, to explore regulatory measures to curb China's use of distillation techniques.

The narrowing gap

Experts now estimate China trails the United States in AI development by just six months. U.S. companies argue that without access to distillation, that gap would be substantially wider, with implications for applications ranging from business planning and drug research to mass surveillance and military weapons systems.

The concern has intensified following the release of GLM-5.2 by Chinese startup Z.ai. The model reportedly rivals top American systems in cybersecurity applications, an area the Trump administration and U.S. AI companies have identified as critical to national security and geopolitical competition.

Why it matters

The distillation controversy highlights a fundamental tension in AI development: the same open access that enables innovation also creates vulnerabilities for technology transfer. If Chinese companies can effectively replicate years of American research investment through systematic querying, it undermines the competitive advantage U.S. firms have built through massive capital expenditures and talent acquisition. The cybersecurity capabilities of models like GLM-5.2 raise particular concerns, as AI systems increasingly play roles in both offensive and defensive cyber operations that could affect critical infrastructure and national security.

The allegations point to a broader challenge facing AI companies as they balance commercial accessibility with protecting proprietary technology from sophisticated copying techniques that operate within technical bounds but outside intended use cases.

These details were first reported by The New York Times.