American AI Under Pressure: China Accused of Massive Copying

Large-Scale Copying Accusations by China

The U.S. government, through Michael Kratsios, the scientific advisor to the Trump administration, has made serious accusations against China. According to a memo written by Kratsios, Chinese actors are using tens of thousands of proxy accounts and jailbreaking techniques to copy American artificial intelligence (AI) models on a large scale.

This process, known as distillation, would allow for the creation of smaller models that are just as effective as American AI systems on specific benchmarks, while significantly reducing development and operational costs. Kratsios criticizes this practice, stating that it brings no innovation and relies on the malicious exploitation of the American industry. Although Kratsios does not mention specific Chinese models, his memo refers to open-source publications and open weights that the U.S. government believes have been built on illegally distilled American models.

Security and Neutrality Issues

Kratsios also emphasizes that during the distillation process, the security protocols of American models are removed. This would compromise the mechanisms designed to keep AI systems neutral and truth-focused. The Trump administration wants American manufacturers to integrate their own political messaging into their systems, similar to China with its domestic models.

While legitimate distillation is an integral part of the AI ecosystem, Kratsios denounces state-sponsored campaigns to steal American research. The administration plans to share intelligence with American companies, strengthen cooperation with the private sector, develop joint countermeasures, and explore actions against those responsible.

Alert from American AI Labs

Companies like Anthropic, OpenAI, and Google have already reported attacks targeting their AI models. The attackers primarily seek to obtain the thought processes of the models, meaning the internal reasoning steps that help tackle complex tasks. These steps are learned through reinforcement learning, a data-intensive and technically complex process.