US Accuses China of Distilling Anthropic’s AI Model – But the Evidence Is Thin

The Trump administration has publicly accused Chinese startup Moonshot AI of stealing technology from Anthropic, alleging that its Kimi K3 model was built by distilling Anthropic’s closed-source Fable model. Michael Kratsios, assistant to the president, stated last week, “We have information indicating that Moonshoot AI distilled Fable from Anthropic for the development of its model K3.” The accusation immediately raised the prospect of sanctions, but a closer look reveals three major gaps in the case.

The context: Kimi K3 has surged in benchmarks, matching performance of Anthropic’s Fable 5 and OpenAI’s GPT-5.6. That success has alarmed Washington, which now frames the incident as a strategic technology theft rather than a routine licensing dispute. Yet the official claim lacks the kind of evidence that would normally accompany such a serious charge.

Evidence gap: claims without data

Kratsios’s statement describes a mechanism: Moonshot AI built a platform to make mass queries to US models, rotating access methods to avoid detection, and transferred those capabilities to Kimi K3. He also alleged that the company accessed servers with Nvidia GB300 chips in countries outside China to evade export controls. But no logs, prompt examples, response patterns, or independent analyses were provided. The accusation is a narrative, not a forensic report. Without reproducible evidence, the claim remains an assertion rather than a demonstrated infringement.

The distillation gray zone

Model distillation is a standard technique in AI: using a larger “teacher” model’s outputs to train a smaller, cheaper “student” model. The industry considers it legitimate when applied to permitted or owned models. The line is crossed when distillation is done on third-party models at scale without permission. The US government argues that is intellectual property theft. However, neither the US nor China has a specific law defining when distillation of a rival model becomes a crime. Violating API terms of service could be a contract breach, but there is no established jurisprudence. Until recently, the US debate centered on regulating frontier models like Anthropic’s Mythos as too dangerous. The rapid pivot to sanctioning distillation shows how fast the legal landscape is shifting without settled rules.

Double standards on training data

The third problem is the most uncomfortable for the US side. American AI companies, including Anthropic, have trained their models on massive datasets that mix websites, books, code, and news articles — much of it copyrighted. Anthropic itself recently reached a legal settlement over copyright claims, and Meta was discovered using terabytes of copyrighted books for training. US companies defend this as “fair use,” but authors neither gave permission nor received compensation. The same industry and government that accepted that practice now condemn a Chinese startup for using outputs of their models to train a rival. While the two practices are not identical — using public data versus using API outputs — the double standard is hard to ignore.

Why this matters for AI developers and users

For developers and businesses relying on frontier models, this accusation signals that the line between legitimate distillation and theft is now a geopolitical weapon. Enforcement without clear law creates uncertainty: any API provider could later claim unauthorized distillation. The case also raises compliance questions for companies using third-party model outputs to fine-tune or distill their own systems. The US is pushing to treat distillation as akin to industrial espionage, a framing that could reshape how models are shared and accessed globally.

What to watch next

No technical evidence has been published, so the next step is whether the US releases any. Kratsios’s mention of Nvidia GB300 chips adds a hardware angle that could be verifiable through supply chain records. The legal framework remains undefined, so any sanctions would be politically driven rather than based on established IP law. Finally, watch how Anthropic responds — the company has not commented publicly on the accusation, and its own fair-use defenses may complicate its position.

Source: Xataka IA – “EEUU está acusando a China de plagiar los modelos de Anthropic. Tenemos tres problemas con esa acusación” by Javier Pastor (https://www.xataka.com/robotica-e-ia/eeuu-esta-acusando-a-china-plagiar-modelos-anthropic-tenemos-tres-problemas-esa-acusacion)