Anthropic’s distilling charges against Chinese firms expose AI training grey area
The US AI lab says DeepSeek, Moonshot AI and MiniMax AI used its systems to improve their models’ capabilities
US artificial intelligence lab Anthropic’s allegation that Chinese AI firms were “distilling” its Claude models has exposed a widely used AI training technique, sparking heated debate over its accepted boundaries, an analyst said.
Anthropic said in a blog post on Monday that DeepSeek, Moonshot AI and MiniMax AI used roughly 24,000 fraudulent accounts to generate more than 16 million exchanges with its Claude models.
The US firm said the activity amounted to unauthorised “distillation” of its systems – a technique that transfers capabilities from a larger model to a smaller one.
Anthropic’s claim reignited accusations against Chinese AI labs over the “distillation” practice - an allegation that has come to the fore as Chinese open-source models have been improving their capabilities rapidly.
In AI development, distillation – formally known as knowledge distillation – is widely used to train smaller models using the outputs of larger systems.
The concept of AI distillation has been around for years. In 2015, three researchers, including Geoffrey Hinton, the “Godfather of AI”, published a paper titled “Distilling the Knowledge in a Neural Network”, showing how knowledge from large models could be transferred to smaller, more easily deployable ones.