Anthropic accuses DeepSeek, other Chinese AI firms of model theft
What's the story
Anthropic, a leading artificial intelligence (AI) company, has accused three Chinese firms—DeepSeek, MiniMax, and Moonshot—of misusing its Claude AI model. The company claims these firms created over 24,000 fake accounts and had more than 16 million exchanges with Claude in an attempt to improve their own products. The allegations were made in an announcement on Monday.
Model manipulation
Distilling AI models
The three companies have been accused of "distilling" Claude, a process where a smaller AI model is trained on a more advanced one. While Anthropic acknowledges distillation as a legitimate training method, it warns that it can also be misused. This includes acquiring powerful capabilities from other labs in less time and at lower costs than developing them independently.
Security concerns
Risks of unprotected capabilities
Anthropic has raised alarms over the potential risks posed by unprotected capabilities. It warns that foreign labs distilling American models could inject these into military, intelligence, and surveillance systems. This could enable authoritarian governments to use advanced AI for offensive cyber operations, disinformation campaigns, and mass surveillance.
Specific allegations
DeepSeek's alleged misuse
DeepSeek, known for its efficient models, allegedly had over 150,000 exchanges with Claude and focused on its reasoning capabilities. The company is also accused of using Claude to create "censorship-safe alternatives to politically sensitive questions about dissidents, party leaders, or authoritarianism." Last week, OpenAI also accused DeepSeek of "ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs."
Additional accusations
Call for action against distillation
Moonshot and MiniMax had over 3.4 million and 13 million exchanges with Claude, respectively. Anthropic has called on other AI industry players, cloud providers, and lawmakers to take action against distillation. The company believes that "restricted chip access" could limit model training and the scale of illicit distillation.