Conditional generative models hold much promise for novel content creation. Whether it is generating a snippet of code, piece of text, or image, such models can potentially save substantial human effort and unlock new capabilities. But there is a fly in this ointment. These models are trained on vast quantities of data, much of which is copyrighted. Due to precedents such as Authors Guild vs Google, many legal scholars believe that training a machine-learning model on copyrighted material constitutes fair use. However, the legal permissibility of using the sampled outputs of such models could be a different matter.

This is not just a theoretical concern. Large models do memorize significant chunks of their training data. For example, if you feed the first sentence of Harry Potter and the Sorcerer’s Stone to GPT-3, it provides the remaining ones: