“The entire English-language section of Wikipedia constituted only 0.3% of the training data for OpenAI's GPT-2 model.”