A recent study published in the journal Patterns reveals that AI image generators tend to revert to a small set of visual styles, regardless of the original prompts given. Researchers tested two models: Stable Diffusion XL for image generation and LLaVA for image description. They conducted an experiment mimicking the game of visual telephone, where an unusual text prompt was converted into an image, described in words, then fed back into the generator to create a new image. This cycle repeated for 100 rounds. The results were striking. Across about 1,000 iterations, the generated images consistently gravitated towards just 12 visual motifs. These included common scenes like lighthouses, Gothic cathedrals, and rainy European cityscapes. The researchers noted that the process often led to these motifs gradually, although some abrupt shifts did occur. Even when the parameters were adjusted or different models were used, the same patterns emerged. By the 100th iteration, most sequences settled on one of the dominant styles, with only minor variations appearing later. Co-author Arend Hintze highlighted the mystery behind occasional shifts from one motif to another, expressing uncertainty about the underlying reasons. The researchers likened the repetitive nature of the results to “visual elevator music,” suggesting that the generated images resemble the generic art often found in hotels or stock photos. Overall, the study shows a clear limitation in the creativity of AI-generated images, revealing a tendency to fall back on familiar themes.