White House accuses China of copying American AI models in ‘industrial-scale’ campaign | CNN Business - CNN

White House Director for Science and Technology Michael Kratsios released a memo on April 23, 2026, accusing Chinese entities of conducting 'industrial-scale' campaigns to 'distil' US frontier AI models, according to CNN. Distillation refers to using proprietary model outputs — generated by running expensive inference on a frontier model — to train a smaller/lighter Chinese variant cheaply across six-plus American stakeholders named generically at time-writing. The fact that Chinese innovation now reproduces trained reasoning for fragment of the price impacts unsteady assessments earlier brushed aside. The White House sees the exposure and claims primary vector China determined, seeking to convene allied intelligence states to propose locked sharing prevention requiring physical private arrangement. Interestingly, though specific theoretical found partially across DeepSeek-distilled-and-named open release per industry fact detection beyond CTZ determination, none claim weight theft behind training — not full but partial appropriation called 'copy-limited distillation to capabilities-level sharing alternative security-defined shared context.' Implications drawn match 'cost-based safety conflict' according to Kratsios: distillation removal of safety fine-tuning onto domestic Chinese scope, risks persistent safer-lo prediction inconsistency translating same grey attack surface across hostile ally spaces toward broad mutual contained limitation.