AIBRIEF
Distillation is a routine AI research technique for shrinking big models into small ones, and the fight now is over who is allowed to do that to whose models.
Distillation is the specific technique that turns a giant, expensive AI model into a small, cheap one. A 'teacher' model generates outputs that a smaller 'student' model is trained on, so the student learns the right behaviours for a task without inheriting the teacher's size, weights, or full capability.
A frontier model that costs billions to train can be shrunk into something a phone, a factory, or a private server can actually run. Newer work passes along the reasoning traces, the steps the teacher used to reach an answer, not just the final answer.
The technique is not new. Researchers have used it for years. The fight now is over consent. US officials and American AI companies have accused Chinese labs of extracting capability from proprietary models without permission, turning a standard research tool into a regulatory flashpoint. Florian Tramèr, an assistant professor at ETH Zurich who studies machine-learning security, has framed the process as closer to a student learning from a book of complicated mathematics, studying the work rather than copying it, per a Reuters explainer republished by LiveMint.
On April 23, 2026, the White House Office of Science and Technology Policy issued a memorandum — NSTM-4, signed by Director Michael Kratsios — committing the administration to information sharing, defensive coordination, and developing best practices against adversarial distillation of US AI models. According to the memo, the US government has information indicating foreign entities principally based in China are engaged in deliberate, industrial-scale campaigns to distill US frontier AI systems, which it characterizes as a threat to US innovation and national security.