on Toshareproject.it - curated by Bruce Sterling
https://www.cell.com/patterns/fulltext/S2666-3899(25)00299-5
The bigger picture
As AI systems increasingly generate and evaluate their own creative outputs, they begin to influence not only what we produce but also how creativity itself evolves. Our study shows that when one combines two state-of-the-art models, one describing images and the other regenerating them, and they interact without human input, they converge toward a small set of highly conventional visual motifs, such as lighthouses, cathedrals, and palatial interiors. This finding reveals that, even without additional training, autonomous AI feedback loops naturally drift toward common attractors—very generic-looking images, which we call “visual elevator music.”
The implication extends far beyond art generation. Many new AI applications use similar self-referential loops. If left unchecked, these systems could amplify the biases and redundancies already present in large datasets, reinforcing aesthetic and cultural uniformity. Understanding how and why such convergence emerges is therefore critical for ensuring that generative AI contributes to diversity rather than homogeneity in culture, design, and knowledge creation. Our work indirectly also addresses the question of large language model (LLM) accuracy, and since these prompt → image → prompt loops drift and converge, we can say with certainty that they are less precise than one would hope. A final conclusion is that human-AI collaboration, rather than fully autonomous creation, may be essential to preserve variety and surprise in the increasingly machine-generated creative landscape.
Highlights
•
AI image-text loops converge to generic motifs despite diverse starting prompts
•
Convergence occurs across all models and temperature settings
•
Systems drift toward “visual elevator music”—stock photography aesthetics
•
Phenomenon mirrors human cultural transmission but lacks corrective feedback…