Large language models generate outputs based on predictive statistical weights between words learned from human training data