Artificial Intelligence image generators often resort to a limited set of visual styles, as revealed by a recent study. Despite having access to extensive databases filled with diverse imagery, these models tend to produce outputs that are strikingly similar and somewhat generic when tasked with evolving prompts. This raises intriguing questions about the nature of creativity in AI.
In a fascinating study published in the journal Patterns (https://www.cell.com/patterns/fulltext/S2666-3899(25)00299-5), researchers examined two prominent AI image generation systems: Stable Diffusion XL and LLaVA. They set up an experimental scenario akin to a visual version of the telephone game. The process unfolded as follows: the Stable Diffusion XL model received a brief, imaginative prompt—such as "While sitting alone amidst nature, I discovered an ancient book containing eight pages written in a forgotten language, yearning to be read and understood." This generated image was then shown to the LLaVA model, which provided a descriptive interpretation. That description was subsequently fed back into the Stable Diffusion model to create another image, and this exchange continued for 100 iterations.
Much like in a classic game of human telephone where the original message can become distorted over time, the initial images produced were quickly transformed. This outcome was not entirely unexpected; anyone familiar with time-lapse videos demonstrating how AI can distort images over successive reproductions (for instance, check out this example: https://www.instagram.com/reel/DSaDQilCKfT/) knows that results can diverge significantly from the original. What did catch the researchers off guard, however, was the revelation that the AI models defaulted to just twelve common visual themes across their output. In their testing of 1,000 different iterations, it became clear that most sequences converged on these dominant motifs, which were largely indistinguishable from one another.
The transition from one image to another typically progressed gradually, though there were instances where the shift occurred abruptly. Researchers expressed disappointment at this finding, coining these prevalent styles as "visual elevator music"—the sort of uninspired art you might encounter in a hotel room. Common scenes included maritime lighthouses, elegant interior settings, bustling urban nightscapes, and charming rustic architecture.
When researchers decided to experiment with alternative models for generating images and descriptions, they discovered similar trends emerging once again. They noted that even when extending the game to 1,000 rounds, the convergence around a specific style still generally took place around the 100th turn, resulting in variations in the subsequent rounds. However, those variations frequently remained tethered to the popular visual motifs already established.
So, what does this all imply? Primarily, it suggests that AI lacks genuine creative prowess. In a human-centered game of telephone, variations can be extreme, as each person's individual biases and interpretations influence the transmission of the message. Conversely, AI appears to struggle with creativity, consistently reverting to a narrow range of styles regardless of how imaginative the initial prompt may be.
It’s important to remember that these AI models are drawing from datasets created by humans. This leads to an interesting consideration about the nature of the data itself and why certain images resonate more with people. Perhaps the key takeaway here is that while replicating styles is relatively straightforward for AI, instilling a sense of artistic taste proves to be a far more challenging endeavor.
What do you think about the limitations of AI creativity? Do you believe it can ever truly replicate human artistic expression, or will it always fall short? Share your thoughts!
