The Trump administration dropped a bombshell Thursday. Foreign actors, mostly from China, stand accused of mounting ‘industrial-scale’ operations to siphon capabilities from America’s leading AI models. Michael Kratsios, director of the White House Office of Science and Technology Policy, laid it out in a memo to federal agencies. ‘The US government has information indicating that foreign entities, principally based in China, are engaged in deliberate, industrial-scale campaigns to distil US frontier AI systems,’ he wrote, as first reported by the Financial Times. That memo, shared on social media, details proxy networks and jailbreak tricks designed to evade safeguards and extract proprietary know-how.
Distillation. That’s the technique at the heart of it. Engineers feed queries into a top-tier model like Anthropic’s Claude or OpenAI’s GPT series, then harvest the outputs to train cheaper knockoffs. The resulting models fall short of the originals’ peak performance. Yet they hit the market fast and affordably, undercutting U.S. innovators. Kratsios called these efforts ‘coordinated campaigns [that] systematically extract capabilities from American AI models, exploiting American expertise and innovation.’ China-based labs don’t match full U.S. prowess. But this shortcut lets them field competitive products without the full R&D grind.
Industry leaders sounded alarms months ago. Anthropic pinpointed campaigns by three China-linked labs, including DeepSeek, trying to ‘illicitly extract Claude’s capabilities,’ per their February statement. OpenAI fired back at DeepSeek too, claiming it aimed to ‘free-ride’ on their work, as noted in a Los Angeles Times report. These aren’t isolated pokes. Reports describe tens of thousands of proxy accounts dodging rate limits and detection, per Reuters.
Timing sharpens the edge. The charges land weeks before President Trump heads to Beijing for a summit with Xi Jinping. Relations already simmer over chip export curbs and trade frictions. Now AI joins the fray. ‘We will be taking action to protect American innovation,’ Kratsios posted on X. Steps include sharing threat intel with AI firms, boosting private-sector coordination, and crafting defenses against these raids. The White House insists legal distillation belongs in open competition. But unauthorized, mass-scale grabs cross a line.
China’s playbook isn’t new. U.S. officials have long flagged IP theft in sectors from semiconductors to aviation. A 2018 arrest of an ex-Apple engineer hauling self-driving tech to Beijing lingers in memory. AI ups the stakes. Frontier models power everything from drug discovery to military simulations. Handing rivals a cheap path erodes the moat U.S. firms built through billions in compute and data.
And the irony bites. U.S. AI giants scrape vast web troves—often copyrighted—for training, arguing fair use. The Trump White House backed that in a March policy framework. Anthropic even bought millions of books, scanned them, then destroyed copies to skirt claims, according to the Washington Post. Yet when Claude’s code leaked, they fired off takedown notices—sparing only a rewritten GitHub version recast in another language. Transformative, they called it. Selective vigilance.
Critics spot hypocrisy. AI firms demand ironclad protection for their outputs while hoovering inputs. The U.S. Copyright Office ruled in 2024 that purely AI-generated works lack human spark for protection, per The New York Times. Distilled models blur lines further. Outputs mimic styles; novels regurgitate verbatim passages, as Ars Technica documented. Who’s stealing from whom?
Beijing stays mum so far. CNN reached the embassy; no reply yet. Past patterns suggest denial or deflection. Meanwhile, export controls bite harder. Trump tightened Nvidia chip flows to China. Labs there scramble for alternatives, fueling distillation drives. A House Select Committee on China eyed this in April, grilling OpenAI, Anthropic, and Google on CCP targeting, as their X post noted.
Markets shrugged. Nvidia dipped slightly; no panic. But insiders watch closely. If distillation scales, it chips at U.S. leads in reasoning and multimodal tasks. DeepSeek’s open-source model already draws eyes for efficiency. U.S. responses could mean tougher API gates, watermarking outputs, or sanctions on offending labs.
Broader fallout looms. Allies like the EU probe AI data practices. Japan and South Korea eye U.S. models warily amid their own builds. Trump’s summit tests resolve. Will he press Xi on theft? Or trade concessions for curbs? History favors blunt talk. First term tariffs reshaped supply chains.
For tech executives, clarity calls. Build better detectors. Share blacklists. The memo pushes that. Kratsios urged agencies to tip off companies on foreign probes. Private detection tools evolve; Anthropic leads there.
America leads the pack. Compute dominance holds. Talent pools deep. But theft tempts copycats. Industrial-scale raids threaten that edge. Trump vows no surrender. Watch Beijing. Watch the models. The race tightens.
WebProNews is an iEntry Publication