New study claims OpenAI may have used copyrighted content in training its artificial intelligence models

The research, conducted by teams from the University of Washington, Stanford and the University of Copenhagen, brings to light a method for detecting whether AI models have 'memorized' parts of their training data - a possible copyright infringement. The study focuses on identifying unique and unusual words in literary texts, known as 'big surprise' words. The results showed that GPT-4, one of OpenAI's models, appeared to have memorized parts of copyrighted fiction books, specifically from a dataset called BookMIA.