Anthropic's policy page permits training classifiers on Claude outputs and prohibits training competing general-purpose models

Sources: primary, discussion
Summary: The page draws the line by what the trained model is rather than by how the outputs were obtained. Sentiment, categorization, summarization, extraction, semantic search and anomaly detection models are permitted. Using outputs as training targets for a general-purpose or open-ended text model is not.
Why it matters: Teams building classifiers on Claude outputs have a stated permission, while teams distilling a general-purpose model do not.