Which technique is used to measure the similarity between the generated output and the ground truth in generative AI models?