If I own Claude's outputs why can't I train my own model on them?

Anthropic's help center article addresses the question of why users who own Claude's outputs cannot necessarily train their own models on them. While customers own the outputs generated from their inputs, the company prohibits using its services to train or develop AI models without written permission, a restriction it says is standard practice across the AI industry.

The restriction exists because Anthropic invests heavily in making Claude safe and helpful, and training on outputs without oversight could lose safety controls, lead to harmful systems, and prevent Anthropic from monitoring deployment. Additionally, using outputs to train competing models would undermine Anthropic's own product offerings.

Permitted uses include training models that do not compete with Anthropic's models, such as specialized classifiers and tools, as well as integrating outputs into applications to power features, generate customer content, analyze data, or improve internal workflows. Prohibited uses are training competitive models and supporting a third party's attempt to do the same.