The method at issue is model distillation: routing a competitor's API at volume, collecting the outputs, and feeding them into your own supervised training pipeline. Each Claude response carries signal about reasoning structure and instruction-following behavior that would otherwise require substantial independent data collection and labeling to replicate. Anthropic says China-based AI labs, including Alibaba and Moonshot AI, did exactly that, unauthorized, at a scale of millions of exchanges.
The mechanism behind the allegation
Distillation from a frontier model works because API access is structurally open. A caller can send programmatic, high-volume queries across many parallel sessions and collect the outputs without triggering obvious anomalies, at least until the volume or patterning of requests exceeds what normal developer or research usage would generate. The responses captured this way carry real training value: they encode how a highly capable model handles complex instructions, which a competing system can learn to replicate without rebuilding the underlying capability from scratch. At the scale of millions of exchanges, that signal becomes a meaningful substitute for original training data.
Anthropic's terms of service prohibit using Claude outputs to train competing AI systems. The company said it detected the unauthorized activity. The disclosure did not specify what enforcement steps, if any, have been taken against Alibaba or Moonshot AI.
The competitive context
Alibaba and Moonshot AI are two of the most active frontier model developers based in China. Alibaba runs the Qwen model family; Moonshot AI operates the Kimi assistant. Both work in the general-purpose instruction-following segment where Claude competes. Funneling millions of Claude exchanges into a training pipeline would, in theory, give either lab a shortcut toward output quality that Anthropic has invested heavily to develop, without bearing the same underlying development cost.
Anthropic has not released a breakdown of exchanges by lab, the period during which the activity occurred, or the detection method used. The company's decision to name Alibaba and Moonshot AI publicly moves the dispute from an internal terms-of-service matter into the open.