A distilled model is trained to approximate another model’s behavior. One route is to expose a high-value model through an API, capture its outputs as labeled data, and train a competing model on them. section 6.6.3
This creates intellectual-property risk even without a direct breach: captured outputs can encode valuable system behavior. Terms of service can restrict scraping, large-scale caching, and competitive reuse of generated outputs, though broad reuse rights may make disputes harder to litigate. section 6.6.3
API outputs as training data
OpenAI alleged that DeepSeek systematically queried ChatGPT outputs to train a competing model. section 6.6.3





