Large language models predict the most probable responses based on internet training data when given generic prompts