When creating images, AI keeps remixing the same 12 stock photo clichés

The Recurring Trap of AI-Generated Imagery: Uncovering the Limitations of Creative Diversity

In the ongoing quest to harness artificial intelligence for creative expression, researchers are increasingly aware of a significant challenge: AI-generated images often fall into repetitive patterns, cycling through a limited set of visual motifs regardless of input diversity. A recent study published in the journal Patterns sheds light on this phenomenon, revealing how AI models tend to remix the same handful of stock photo clichés—what some are calling “visual elevator music.”

The Visual Telephone Analogy

To understand this phenomenon, it helps to consider a playful analogy known as the “visual telephone” game. In this game, one player sketches an image and describes it verbally to another, who then attempts to recreate the drawing based solely on that description. This iterative process highlights how information can distort or converge over exchanges, and it mirrors how AI models generate images through a sequence of transformations and interpretations.

The Study: Pairing AI for Repetitive Results

In the recent study, researchers paired two AI models—one responsible for generating images and the other for interpreting prompts— and subjected them to 100 rounds of this “visual telephone” process. They began with a broad spectrum of prompts, ranging from abstract concepts to specific scenes, expecting a diversity of outputs to emerge over multiple iterations.

However, the results were surprising. Regardless of how varied or nuanced the initial prompts, the AI models consistently converged on a set of just 12 visual motifs. These motifs, frequently Eurocentric in perspective, included clichés like serene landscapes, stereotypical cityscapes, and generic portraits—elements that resemble “stock photos” more than original art. The researchers described this pattern as “visual elevator music,” highlighting its monotonous, repetitive nature.

Implications for AI Creativity and Bias

This tendency raises important questions about the creative potential and limitations of current AI image generation technologies. When diverse prompts lead to a narrow set of visual outputs, it suggests that these models rely heavily on existing, often stereotypical, visual data they’ve been trained on. This not only stifles genuine creativity but also perpetuates cultural biases embedded within the training datasets.

Furthermore, the Eurocentric bias identified in these recurring motifs underscores the necessity for more diverse and inclusive training data to foster richer, more varied AI-generated imagery.

Moving Toward Greater Diversity in AI Art

As developers and users of AI art tools, recognizing this pattern is the first step toward addressing it