AI company Anthropic has requested a US government investigation, claiming that Chinese AI firms, including Alibaba, have illicitly extracted the intelligence of its 'Claude' model.
Imagine this: You have spent years staying up late to create a very smart and creative personal tutor. Then, one day, someone secretly records and compiles all of your tutor’s lessons, creates an identical “fake tutor,” and starts offering lessons at a dirt-cheap price. How would you feel?
Recently, a claim has surfaced that exactly this kind of thing is happening in the artificial intelligence (AI) industry. Anthropic, the developer of the AI model ‘Claude,’ has accused several Chinese AI companies, including Alibaba, of secretly stealing the intelligence of its model. Source 1
Why does this matter?
This incident goes beyond a mere corporate dispute and raises important questions about how the “intelligence of AI”—which has become deeply embedded in our daily lives—is created and protected. In reality, creating a single AI model costs astronomical sums and requires the efforts of countless researchers. Source 9 If someone can clone this immense effort for a bargain, the incentive for companies to develop new AI will be severely dampened. This could ultimately slow down the pace of technological development and harm fair market competition. Source 8
Easy to understand: Theft by ‘Distillation’
Anthropic calls this incident a ‘Distillation’ attack. Source 17 Distillation is a common training method in the AI industry; to use an analogy, it’s like secretly copying a “summary cheat sheet” from a brilliant teacher.
Originally, distillation (a technique where a smaller model learns by following the answers of a larger model) is a legitimate training method used to make a lower-performing “student AI” smarter by referencing the answers of a high-performing “teacher AI.” Source 6 However, what Anthropic is taking issue with is that this process was carried out on a massive scale “without permission.”
It’s as if a chef secretly broke into a kitchen to figure out how to make a secret sauce and tested the ingredient ratios 28.8 million times. Source 8 Anthropic claims that a specific research team at Alibaba created as many as 25,000 fake accounts and bombarded Claude with questions in an attempt to extract this secret recipe. Source 8, Source 13
Where we stand: Anthropic’s counterattack
Anthropic is taking this matter very seriously. Going beyond mere protest, they have sent official letters to US Senators and White House officials to notify them of this situation. Source 2
It is not just Alibaba that has been named. Anthropic revealed that a total of three Chinese AI developers—DeepSeek, Moonshot AI, and MiniMax—illegally extracted their technology in a similar manner to improve their own models. Source 3, Source 5 This issue has now become a hot potato, intertwined with international competition for AI technological hegemony. Source 16
What comes next?
This incident is likely to spark new legal discussions about the ‘copyright of AI intelligence.’ While the focus until now has been on the copyright of software or content, protecting the “essence of intelligence” that an AI model has learned will become a core challenge.
For users, it remains to be seen whether these incidents will lead to AI developers strengthening their security policies, resulting in a safer AI environment, or if it will result in inconveniences like service restrictions. One thing is certain: as AI becomes smarter, the tension between attempts to steal it and efforts to protect it will only increase.
References
- Anthropic says Alibaba illicitly extracted Claude AI model capabilities
- Anthropic Accuses Alibaba Of Running Major Adversarial Distillation Campaign
- China’s AI Companies Illicitly Extract Claude Capabilities
- Anthropic accuses Chinese AI firms of siphoning Claude via distillation
- Anthropic Says Chinese AI Companies Improved Models By ‘Illicitly Copying’
- Anthropic accuses Chinese labs of stealing Claude’s data
- Anthropic alleges large-scale distillation campaigns targeting Claude
- Alibaba Illicit AI Model Extraction Accusations Cause Shares to Drop
- Claude
- Anthropic Accuses Chinese AI Firms of Illicit Model “Distillation”
- Anthropic Writes To White House Accusing Alibaba Of “Illicitly” Accessing Claude AI
- Anthropic says Alibaba illicitly extracted Claude AI model capabilities
- Anthropic says Alibaba illicitly extracted Claude AI model capabilities
- Anthropic Accuses Alibaba of Distilling Claude AI Model Capabilities
- Directly deleting AI data
- Using the answers of a powerful AI as training data to copy its performance
- Physically hacking AI servers
- About 250
- About 2,500
- About 25,000
- United States
- China
- South Korea