“Comparing AI models based on a single response to a single prompt is a flawed methodology because it fails to account for the distribution of possible outcomes.”