U.S. orders global warning on AI theft

The U.S. State Department has ordered embassies and consulates worldwide to warn governments about alleged Chinese theft of American artificial intelligence models. (reuters.com) (cnbc.com) A diplomatic cable dated Friday said the campaign should highlight what Washington called efforts by Chinese firms, including DeepSeek, Moonshot AI and MiniMax, to steal intellectual property from U.S. AI labs. (reuters.com) (cnbc.com) The cable told diplomats to speak with foreign counterparts about “extraction and distillation” of U.S. models. In plain terms, distillation means training a smaller system on the answers produced by a larger, more expensive model. (reuters.com) (pbs.org) Washington says that process can copy useful capabilities without paying the original developer. The cable said distilled systems can look competitive on some benchmarks while omitting the full performance and safety controls of the original model. (reuters.com) (cnbc.com) The warning landed the same week Michael Kratsios, President Donald Trump’s science adviser, accused foreign entities “principally based in China” of running industrial-scale campaigns to extract capabilities from leading U.S. systems. (pbs.org) (abcnews.go.com) Kratsios said the administration would work with American AI companies to identify the activity, strengthen defenses and find ways to punish offenders. The House Foreign Affairs Committee also backed a bill this week to identify foreign actors extracting key technical features from closed-source U.S. models and sanction them. (pbs.org) (abcnews.go.com) China rejected the accusations. Chinese Embassy spokesperson Liu Pengyu called them “baseless allegations,” and Foreign Ministry spokesperson Guo Jiakun said Washington was smearing China’s AI industry and should “respect facts.” (pbs.org) The timing also reflects how much the AI race has tightened. Stanford’s 2026 AI Index said the U.S.-China frontier model gap has “effectively closed,” with Anthropic’s top model leading the top Chinese model by 2.7% as of March 2026. (hai.stanford.edu) (pbs.org) DeepSeek underscored that pressure on Friday by previewing a new model adapted for Huawei chips, a sign Chinese developers are building around domestic hardware as U.S. restrictions tighten. (reuters.com) (cnbc.com) The cable said a separate demarche had been sent to Beijing, and it framed the outreach as groundwork for possible follow-up action by the U.S. government. The next step is no longer just export controls on chips; it is a diplomatic campaign over how AI models themselves are copied. (reuters.com) (cnbc.com)