The US government says China is systematically copying American AI models at massive scale.
According to a memo from Michael Kratsios, President Trump’s science advisor, foreign actors are using tens of thousands of proxy accounts and jailbreaking techniques to pull capabilities from top US models. The process, known as distillation, creates smaller models that match the performance of large US AI systems on select benchmarks, but at a fraction of the development and operating cost.
There is nothing innovative about systematically extracting and copying the innovations of American industry, and there is nothing open about supposedly open models that are derived from acts of malicious exploitation.
Michael Kratsios
Kratsios doesn’t name specific Chinese models, but his memo points to open-source and open-weight releases that the US government believes were built on illegally distilled American frontier models.
Safety guardrails stripped during the copying process
Kratsios says the actors also strip safety protocols from the resulting models during distillation. In his view, that undermines mechanisms designed to keep AI systems “ideologically neutral” and “truth-seeking.” The Trump administration wants US model makers to bake its own political messaging into their systems under the banner of neutrality, much like the Chinese government does with its domestic models.
Kratsios stresses that legitimate distillation is a valid part of the AI ecosystem. But systematic, state-backed campaigns to steal American research are off the table, he says. The Trump administration now plans to share intelligence about these campaigns with US companies, tighten cooperation across the private sector, develop joint countermeasures, and explore action against those responsible.
US AI labs have been raising alarms for a while
Anthropic, OpenAI, and Google have all flagged attacks on their AI models before. The attackers are mostly after the models’ chains of thought, the internal reasoning steps that help models work through more complex tasks. These are learned through reinforcement learning, a technically demanding and data-intensive process.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive “AI Radar” frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
