Boyi Wei 1 Yangsibo Huang 1 Noah A. Smith

These models can memorize and generate content similar to their training data, posing potential concerns. Therefore, model creators are motivated to develop mitigation methods that prevent generating protected content.