The battle over intellectual property in artificial intelligence has reached a new level of tension. Anthropic, the company founded by former OpenAI employees, has filed a complaint with the U.S. government alleging a massive campaign conducted by groups linked to Alibaba and its Qwen lab. According to the letter sent to lawmakers, nearly 25,000 fake accounts were used to generate over 28.8 million interactions with the Claude model, aiming to extract its most advanced capabilities through a technique known as model distillation.
How illicit model distillation works
Model distillation is a common practice in the industry, used to create smaller, faster versions of a larger model. However, when employed by a competitor without permission, it becomes a potential intellectual property violation. Anthropic's accusation claims that the fraudulent queries were highly specific and complex, focusing on advanced software engineering and agentic reasoning capabilities. Each response from Claude helped reconstruct the model's behavior, accelerating the development of competing systems by Alibaba. The vulnerability is inherent: large language models are designed to answer, and every answer reveals information about their inner workings. At a scale of tens of millions of questions, conversation becomes reverse engineering.
Sponsored Protocol
The irony of the AI arms race
This is not the first time Anthropic has raised similar allegations; it previously targeted DeepSeek, Moonshot AI, and MiniMax. OpenAI has also expressed concern about being victimized by the same technique. The irony is hard to ignore: the very companies that collected vast amounts of publicly available data, including copyrighted materials, to train their models now fiercely defend their models' behavior as intellectual property. If a model can be imitated simply by asking the right questions, the billion-dollar investment in building frontier models becomes questionable.
Sponsored Protocol
Implications for innovation in the field
Anthropic has urged lawmakers to act quickly to stop this practice. The fear is that without adequate protection, there will be little incentive to innovate: anyone could copy years of research at a negligible cost. Competition would devolve into a race against clones, making it hard to distinguish a model born from original research from a replica. As the AI community awaits Alibaba's response, the future challenge may not be building the smartest model, but preventing others from learning how it operates, one question at a time.
Sponsored Protocol
For further reading, check Anthropic Regains Partial Access to Mythos Model and OpenAI Launches GPT-5.6 Sol, Terra and Luna in Limited Preview. For broader context, see the Wikipedia entry on model distillation.
