Something odd started showing up in public conversations with DeepSeek, a Chinese AI model. Users would ask which model they were talking to, and DeepSeek would answer as if it were Claude, the AI assistant made by Anthropic. A public bug report on GitHub, the code-sharing site, documented the pattern. On its own, that does not prove theft, since similar mix-ups have been seen in other models. Still, it is a visible clue that another model’s answers may have found their way into DeepSeek’s training.
Anthropic later made the concern far more serious in a public report. It says it identified industrial-scale campaigns by DeepSeek and two other Chinese AI companies, Moonshot and MiniMax, to illicitly extract Claude’s capabilities. According to the report, the campaigns generated more than 16 million exchanges with Claude through about 24,000 fraudulent accounts. They targeted Claude’s strongest skills, including reasoning, coding, data analysis and the use of software tools.
The method at issue is training a smaller model on the answers of a stronger one, known as distillation. In normal research, it is a legitimate way to make models smaller, cheaper and easier to run. AI researchers Geoffrey Hinton, Oriol Vinyals and Jeff Dean described it as a way to compress the knowledge in large systems into smaller ones. Inside one company, that is an efficiency tool. Across company lines and without permission, it becomes a shortcut for copying years of work.
That shortcut skips a very large bill. Building a strong model from scratch takes huge amounts of data, long development cycles and major spending on computing power. Then comes a long phase of tuning, testing and refinement before the model is useful, reliable and safe enough to release widely. One company pays for all of that work. Another may try to learn from the finished answers, which Anthropic says gives competitors powerful capabilities in a fraction of the time and cost of building them.
Cheap copying shortens the period in which the original builder benefits from paying for the work. Protecting those capabilities is becoming part of the cost of staying ahead. Anthropic says it is responding with stronger detection, tighter verification and countermeasures designed to make large-scale extraction harder and less useful.
