Semantic Alignment is Not Enough: Evaluating Creativity in Iterative AI Image Generation

As Generative AI systems are increasingly used as creative partners rather than tools, aligning computational and human evaluations of creativity has become increasingly important. Using a Telephone game-inspired setup, this work examines how co-evolutionary generative processes shape semantic alignment and perceived creativity. Specifically, we used an iterative text-to-image pipeline where semantic alignment was measured using CLIPScore. We systematically varied prompt design (length and fictionality) and model variance. Human participants evaluated generated images on perceived creativity, novelty, surprise, value, and prompt fit, enabling a comparison between subjective judgments and an automated measure. CLIPScore showed a persistent decline across iterations, indicating semantic drift with diminishing changes after approximately ten iterations. While semantic alignment was a significant positive predictor of human creativity ratings, more complex prompt conditions revealed differences between automated metrics and subjective judgments. Our findings suggest that creative judgment involves more than semantic alignment, incorporating subjective interpretation and contextual information not fully captured by the computational measure. We argue that automated similarity metrics should be treated as complementary rather than proxies for creativity, and that new evaluation frameworks that integrate both subjective and computational evaluations of creativity are required to align human and machine perspectives on creative quality.