The US is accusing China of plagiarizing Anthropic models. We have three problems with that accusation.
Michael Kratsios, assistant to the president of the United States, did not bite his tongue this week when claimed that “We have information indicating that Moonshoot AI distilled Fable from Anthropic for the development of its K3 model.” Or in other words, they accuse China of plagiarizing the American company’s advanced AI model. There are several fundamental problems with that accusation. Let’s see them. Fear of Kimi K3. This model has just burst onto the AI scene with overwhelming force. Benchmarks show that Kimi K3 is one of the best AI models in the worldwith performance close to Fable 5 and GPT-5.6, the best public models from Anthropic and OpenAI. This milestone has triggered its popularity among users and companies, and also the alarms in Washington. The immediate consequence has in fact been political, because since the Trump Administration there is now open talk of sanctions for “theft of technology” if the accusation is confirmed. The accusation. In his text, Kratsios explains the supposed mechanism used by Moonshot AI to copy his model. It states that the Chinese startup would have built an internal platform to make mass queries to US models, changing access methods to avoid detection and then transferring its capabilities to Kimi K3. He also adds that the company has managed to access servers with Nvidia GB300 chips in countries outside China to avoid export restrictions. Where is the evidence? The first problem we have with that speech is that a key piece of that speech is missing from the start: Kratsios does not support this accusation with technical evidence. It does not provide usage records, it does not provide examples of prompts and responses, nor query patterns. Nor does it provide independent analyzes that any researcher could review. What we have right now is an official story and accusation, not a proven “infringement.” The controversy over distillation. The key word in that speech is “distillation.” In AI, distill a model It means training a new one using the responses of another. The “master” model answers thousands or millions of questions, and the “student” model learns to imitate its reasoning patterns at lower cost and size. The industry itself describes distillation as a legitimate and standard technique when applied to permitted or proprietary models. The red line appears when it is applied to third-party models, on a large scale and without permission, which according to the US Government is theft of intellectual property. But crime, what is called crime, is not. The second problem is that the statement and the accusation are not supported by any clear and defined legal framework. Neither the US nor China currently have a specific law that says under what conditions distilling a rival model is an intellectual property crime. You may violate an API’s terms of service or even end up obtaining sensitive information from companies, but there is no jurisprudence about it. In fact, until recently the debate in the US was about regulating US frontier models like Mythos, which were considered too dangerous. Suddenly the debate is now on sanctioning the distillation of models. Everything is moving too quickly, once again, to adapt the legislation. Hypocrisy made in USA. The third big problem we have with that accusation is that the big American models—including those at Anthropic—have been trained on massive data sets that mix websites, books, code, newspaper articles, and other materials of which a notable portion is protected by copyright. own Anthropic reached an agreement with justice these days for that reason, Meta has already discovered it stealing terabytes of copyrighted books to train your models. Double yardstick. In the US, AI companies defend themselves by arguing that That falls under “fair use” of the content, but the authors neither gave their permission nor charged for it. That same industry and that same Government that have made it normal to train models on other people’s content complain that a Chinese startup uses the outputs of its models to train its own. They are not identical practices, but the double standard is difficult to ignore. AI, once again, as a weapon. The Trump administration is taking this issue to another scale. By accusing Moonshot AI of using restricted chips and developing Kimi K3 by distilling Fable 5, he places distillation in the same box as industrial espionage or the theft of military secrets. AI is no longer a software product and becomes a strategic technological weapon. And there it is not so important to provide evidence or be technically precise: it is important that what China has done is an attack on its (former?) technological hegemony. If this type of accusation sounds familiar to you, you have a good memory. In Xataka | An AI model did not have access to the internet. So he thought it was better to have it and decided to hack something along the way
