When Anthropic accused Alibaba of "illicitly extracting" Claude's capabilities, they were referring to a practice known as model distillation. This is where you use a powerful model's outputs — its answers, reasoning steps, and examples — as training data to teach a smaller or different model to behave similarly.

Knowledge distillation is actually a legitimate and common technique in AI, used to compress large models into smaller, faster ones. The controversy arises when it's done without permission, essentially using a proprietary model's intelligence to bootstrap a competitor. Most major AI providers explicitly forbid using their model outputs to train other models in their terms of service.

The deeper issue is that a model's "capabilities" aren't easily copyrighted like code or text. This makes model theft a murky legal and ethical frontier — one that's becoming increasingly important as the gap between frontier models and open-source alternatives narrows.