The White House has accused Chinese AI company Moonshot of building its powerful Kimi K3 large language model by copying Anthropic’s Fable model through a process known as distillation, while using chips barred from export to China. Science adviser Michael Kratsios said in a social media post that the alleged theft of proprietary US technology was “unacceptable”, and Treasury Secretary Scott Bessent claimed “watermarks” of US models had been found in Chinese counterparts. Neither official provided evidence, and Moonshot did not respond to requests for comment.
However, AI researchers are pushing back against the narrative. Braden Hancock of the Laude Institute and Snorkel AI told TechCrunch that the timeline alone undermines the distillation claim: Fable was only publicly released on 1 July 2026, and Kimi K3 appeared just two weeks later. “You can’t distill that much data, train a model, and release it in two weeks,” he said. Nathan Lambert of the Allen Institute for AI added that distillation via supervised fine-tuning is becoming less effective as frontier models grow more complex, and that replicating Fable’s capabilities would likely require reinforcement learning at a scale that is prohibitively expensive and slow.
Distillation typically involves querying a target model to generate training data, sometimes extracting its chain-of-thought reasoning. Lambert noted that while earlier Chinese models may have used distillation from Anthropic’s previous models — Anthropic earlier this year accused Moonshot, DeepSeek and MiniMax of systematic extraction — the leap to Fable-level performance in such a short window is not explained by that technique alone. He argued that the benefits of supervised fine-tuning are diminishing, and that genuine frontier-level distillation would demand tens of millions of reinforcement learning agents running against an API, which would be both costly and time-consuming.
The allegations come amid reported discussions in Washington about banning Chinese open-weight models, which has unsettled the global AI sector. For UK businesses and consumers, the row underscores the growing geopolitical friction around AI. The UK’s Information Commissioner’s Office has previously warned that AI firms must comply with data protection law when training models, while the EU AI Act imposes transparency obligations on developers. If the US tightens export controls or pushes for a ban on Chinese open-weight models, UK companies that rely on such models for research or commercial applications could face disruption.
Experts caution that the focus on distillation may distract from the real story: Chinese AI labs are rapidly closing the gap through legitimate innovation. “I don’t think you get a model this strong and this quickly on the heels of Fable doing strictly distillation,” Hancock said. The UK, which has positioned itself as a hub for AI safety and innovation, will need to navigate these tensions carefully. The ICO has not yet commented on the specific allegations, but the broader regulatory landscape is likely to evolve as the US-China tech rivalry intensifies.