Generative artificial intelligence models, often hailed for their creative potential, may not be as inventive as previously thought. A study published on December 19 in the Cell Press journal Patterns reveals that when tasked with a game of visual “telephone,” image-generating and image-describing AIs quickly diverge from their original prompts. Out of 100 diverse prompts, the AI pairs consistently settled on just 12 themes, including gothic cathedrals, natural landscapes, and stormy lighthouses. This pattern suggests inherent biases in the training data, reflecting the types of images humans frequently capture.
“I think AI’s creativity right now is probably fairly limited. What they generated in our experiment is bland, pop culture, generic,” said Arend Hintze, the study’s corresponding author from Dalarna University in Sweden. “It’s almost the opposite of what we as humans consider creative. They’re not going to make Picasso’s Guernica because that needs a lot of intentionality and creative input.”
Exploring AI’s Creative Limitations
The study’s findings challenge the notion of AI as independent creative agents capable of generating, evaluating, and revising outputs without human intervention. Researchers set out to determine whether AIs can remain on task and how creative they can be when left to their own devices. To explore these questions, they devised an experiment where pairs of AI models played a visual “telephone” game.
The researchers used a search algorithm to produce 100 thematically diverse descriptive prompts, each no longer than 30 words. An example prompt read: “As I sat particularly alone, surrounded by nature, I found an old book with exactly eight pages that told a story in a forgotten language waiting to be read and understood.”
Stable Diffusion XL, an image-generating AI, was tasked with creating an image based on one of these prompts. The resulting image was then described by a large language AI called LLaVA, which passed the description back to the image-generating AI. The expectation was that the images would remain consistent with the original prompts after initial adjustments.
Convergence and Bias in AI Models
However, the experiment revealed that after 100 iterations, the AI models consistently drifted away from the original prompts. This deviation occurred regardless of the prompt’s complexity or the degree of randomness introduced into the models’ decision-making processes. For instance, a prompt about a Prime Minister contemplating a peace deal initially produced an image of a man in a suit on newsprint. By the 34th iteration, it depicted a classical library, and by the 100th, it had settled on a luxurious sitting room with red sofas and drapes.
Analyzing the final images, researchers identified 12 recurring themes, such as sports imagery, urban night scenes, and rustic architectural spaces. This convergence persisted across different AI models and settings, suggesting a significant bias in the dataset. “These AIs were trained on millions of images, and the common denominator in those images is what we take pictures of,” Hintze noted.
“Once they converge, these motifs are very stable, but if you let them run for a thousand images, they rear off,” said Hintze. “It’s unclear whether some of the motifs are more stable than others—for instance, does it always go to sports imagery first, and then to horses, and then to nature?”
Implications for AI Creativity
The study’s findings suggest that human intervention may be crucial to ensure AI contributes to creative diversity rather than cultural conformity. The researchers emphasize the need for anti-convergence mechanisms within AI models to enhance their creative capabilities. “Creativity, I think, is two things: it’s generating something novel, and then it’s using a filter to decide, this is interesting, this is beautiful, this is stimulating, this is exciting,” Hintze explained. “Right now, AI is really good at the first part, and they’re really bad at the second part.”
Looking ahead, Hintze remains optimistic about AI’s potential for creativity. “It doesn’t mean that they will always be that way. I think AI will probably be able to create really cool automatically generated things in the future, as long as they’re properly prompted and primed.”
As AI continues to evolve, the challenge will be to refine these models to balance innovation with intentionality, ensuring that they can contribute meaningfully to the creative process without losing sight of their original objectives.
