The White House has leveled a blunt accusation: Chinese-based actors are running an “industrial-scale” campaign to copy U.S. frontier AI models.
What happened
On April 23, Michael Kratsios, director of the White House Office of Science and Technology Policy, circulated a memo to U.S. government agencies alleging that groups “principally based in China” are conducting deliberate, large-scale efforts to distill and steal cutting‑edge American AI systems. The memo — the most explicit U.S. government statement to date about alleged Chinese IP theft in AI — arrived just days before a scheduled Trump–Xi summit.
How the alleged campaign works
The memo describes a multi-step operation:
- Distillation: actors train smaller models on the outputs of larger proprietary systems, letting them approximate frontier performance without the huge costs of training from scratch.
- Scale and evasion: tens of thousands of proxy accounts are used to flood U.S. models with queries, bypassing rate limits and detection systems designed to spot suspicious usage.
- Jailbreaking: techniques are applied to strip safety filters and reveal proprietary model behavior, which is then captured as training data for distilled copies.
Who’s implicated and what’s already happened
OpenAI and Anthropic have previously accused a Chinese company called DeepSeek of distilling their models. The day after Kratsios’ memo surfaced, DeepSeek released preview versions of a V4 model (April 24) intended to run on Huawei’s Ascend chips — a move some Chinese rivals viewed as a competitive threat: shares of Zhipu AI and MiniMax fell about 9% and 7% respectively on V4’s launch day. DeepSeek denies illegal use of synthetic training data, saying its dataset comes from organic web searches.
Reaction and fallout
The Chinese Embassy in Washington called the accusations “baseless” and said Beijing “attaches great importance to the protection of intellectual property rights,” while the Foreign Ministry urged Washington to “abandon biases.” U.S. AI firms including Nvidia, Microsoft and Meta have been navigating growing geopolitical exposure to China through 2026, and an earlier April development — the IRGC designation of American tech companies as targets — added another layer of pressure.
U.S. response and policy levers
The memo says the administration will begin sharing intelligence about active distillation campaigns directly with American AI companies and will “explore a range of measures to hold foreign actors accountable.” Separately, Commerce Secretary Howard Lutnick confirmed on April 23 that, despite a conditional approval in January, no advanced Nvidia AI chips had actually been shipped to China — injecting a hardware and supply-chain dimension into an already tense dispute over models and data.
Why it matters now
Industry executives are closely watching whether AI intellectual property and chip export issues will be raised at the upcoming Trump–Xi summit. Nvidia’s revenue has already been materially impacted by tightening export restrictions that accelerated through 2025 and into 2026, so any diplomatic outcome on AI trade could have major consequences for the U.S. semiconductor industry and the broader AI ecosystem.
What the memo didn’t do
The White House memo did not name specific Chinese companies or individuals, nor did it announce immediate sanctions or penalties. It signals, however, a firmer willingness by the administration to share intelligence with industry and consider accountability measures — and it puts AI model theft squarely on the bilateral agenda as U.S.–China tensions over technology continue to escalate.
Read more AI-generated news on: undefined/news
