The White House is accusing China of running a coordinated, large-scale campaign to steal proprietary knowledge from leading American AI companies. The accusation comes from a memo written by Michael Kratsios, director of the White House Office of Science and Technology Policy, seen by the Financial Times.
The memo's language leaves little room for interpretation: "foreign entities, principally based in China, are engaged in deliberate, industrial-scale campaigns to distil US frontier AI systems." The timing matters. It arrives ahead of a planned meeting between President Trump and Chinese President Xi Jinping in Beijing, where technology competition is expected to dominate the agenda.
How it works
The technique at the center of the allegations is called model distillation: systematically querying an AI system to extract its reasoning patterns, then using that data to train a cheaper, competing model. The method has been publicly documented for years. What is new, according to the White House, is the scale. Kratsios says foreign actors are using tens of thousands of proxy accounts to avoid detection, along with jailbreaking techniques to extract information that systems would otherwise decline to share.
The challenge for AI companies is genuine. Distinguishing a legitimate user from someone methodically probing a model to replicate it is genuinely difficult. By the time the pattern becomes detectable, a significant amount of extraction has already happened.
What the government is doing about it
The administration says it will share intelligence with American AI firms to help them identify these attempts. That is a cooperative posture rather than a regulatory one, though probably not the last word. The US House is considering legislation that would penalize AI model copying by Chinese companies. A Senate Judiciary Committee hearing on Chinese intellectual property theft last week also spent considerable time on AI specifically.
The bigger picture
Something worth sitting with: Stanford's 2026 AI Index found that China has closed the performance gap with American frontier models to just 2.7 percentage points, down from as much as 31 points in 2023. China managed this while spending roughly 23 times less on private AI investment than the United States.
That raises a real question about what the alleged theft is actually buying. Organic research progress and stolen knowledge are not mutually exclusive. If the distillation campaigns are as large as the memo describes, they could meaningfully shorten development timelines for a research establishment already advancing quickly on its own. For AI companies, the immediate takeaway is practical: rate-limiting suspicious query patterns and monitoring for distillation behavior are no longer optional exercises.
The accusation also represents a significant escalation in how the US government is framing its relationship with China on AI. Moving from "rivalry" to "theft" carries real diplomatic weight, and that framing will be impossible to ignore when Trump and Xi sit down in Beijing.
Sources
- i. www.usnews.com
- ii. invezz.com
- iii. www.bloomberg.com
- iv. ipwatchdog.com
Commentarii · 0