The White House Accuses China of Technology Theft through AI

Tensions between the United States and China are escalating over artificial intelligence, as the White House accuses Beijing of industrial distillation practices. This accusation comes three weeks ahead of the Trump-Xi summit scheduled for May 14 in Beijing.

From Private Complaints to a National Security Issue

American companies OpenAI and Anthropic initially raised these concerns in February, accusing Chinese entities of distilling their AI models. OpenAI identified accounts linked to DeepSeek using proxies to mask the origin of requests, while Anthropic reported 16 million fraudulent exchanges with its model Claude. On April 23, the White House formalized these accusations through a memo signed by Michael Kratsios, director of the OSTP.

This document, referenced NSTM-4, is addressed to all federal departments and agencies, highlighting the threat these practices pose to national security. Although the memo does not explicitly name Chinese companies, it mentions "foreign entities, primarily in China."

Enhanced Cooperation but No Immediate Sanctions

Washington has decided to share intelligence with American AI companies to counter these distillation attempts, drawing inspiration from cybersecurity practices. OpenAI, Anthropic, and Google had already begun pooling their alerts through the Frontier Model Forum. However, the memo remains a policy document without enforceable power, as no sanctions or coercive measures have been announced.

Diplomatic and Legislative Context

The publication of this memo comes amid a tense diplomatic backdrop as the Trump-Xi summit approaches. Distillation adds to American grievances regarding chip exports and semiconductor counterfeiting. Meanwhile, Congress is moving forward with the introduction of the "Deterring American AI Model Theft Act" by Representative Bill Huizenga, aimed at sanctioning entities engaged in unauthorized distillation.

The current legal framework does not clearly cover distillation, which involves training a smaller model on the responses of a larger model. While Apple has a contractual right to do this with Google’s Gemini, unauthorized access via fraudulent accounts remains problematic. Washington hopes that the public designation of this threat will be sufficient to influence upcoming negotiations.