OpenAI's models 'memorized' copyrighted content, new study suggests