US Accuses Chinese AI Firm of Data Scraping

Michael Kratsios, a former technology advisor to the Trump administration, has accused Chinese AI startup Moonshot AI of intellectual property theft. According to Kratsios, Moonshot AI extensively scraped U.S. models, particularly Anthropic's leading Fable model, to develop its new Kimi K3 language model. He alleges the company utilized a "sophisticated internal platform" enabling rapid switching between multiple access methods to evade detection.

The U.S. government is increasing scrutiny of Chinese AI models. U.S. Treasury Secretary Scott Bessent announced that Chinese "open-weight" models will be investigated for intellectual property theft, with potential sanctions threatened. This move reflects growing concerns in Washington about China's advancements in AI technology.

Within the U.S., policymakers are divided on how to handle China's openly available AI models. Models like Kimi K3, which offer open data, are attractive to developers but pose challenges for regulation. A coalition of nearly 200 Silicon Valley companies has urged the Trump administration against broad bans, warning that such measures could harm U.S. startups while doing little to halt the global spread of these models.

OpenAI President Greg Brockman acknowledged Kimi K3 as a "pretty good model" but expressed uncertainty about whether Moonshot AI unlawfully distilled OpenAI's outputs. Model distillation, where one AI is trained on the outputs of another, is a common practice. However, concerns are rising about the ethical boundaries and potential misuse of this technique, particularly when proprietary data may be involved.