AI image generators converge on just 12 visual clichés, study reveals creative limitations

AI Image Generators Trapped in Creative Loop

A study published in Patterns has exposed a troubling limitation in AI creativity: when left to generate images autonomously, AI image generators consistently converge on a limited set of 12 visual styles, regardless of how diverse or unusual the initial prompt 1

1

. Researchers at Dalarna University designed a visual telephone experiment where they paired Stable Diffusion XL with LLaVA, an image-describing model, and set them loose for 100 rounds of iterative generation 2

2

.

The results reveal a pattern that challenges assumptions about generative AI capabilities. Despite starting with 100 deliberately unusual AI prompts—one reading "As the morning sun rises over the nation, eight weary travelers prepare to embark on a plan that will seem impossible to achieve but promises to take them beyond"—the systems repeatedly drifted toward the same generic themes 1

1

. Study co-author Arend Hintze notes, "You cannot get [the prompts] further away from each other. We tried to make them as wild as possible."

Visual Clichés Dominate Output

The convergence on visual motifs happened quickly and predictably. A prompt about a prime minister grappling with a fragile peace deal devolved into an image of a pompous sitting room with a dramatic chandelier after just a few dozen handoffs 1

1

. Other trajectories regularly drifted toward Gothic cathedrals, pastoral landscapes, rainy nighttime scenes in Paris, maritime lighthouses, formal interiors, urban night settings, and rustic architecture 2

2

.

Researchers dubbed these recurring themes "visual elevator music"—the type of meaningless, broadly acceptable stock imagery you might find in Ikea picture frames or hotel rooms 1

1

3

3

. The trend persisted even when researchers adjusted randomness parameters and swapped in different AI models, demonstrating that this limitation reflects fundamental constraints rather than quirks of specific systems.

Training Data Bias Drives Homogenization

Ahmed Elgammal, director of the Art and Artificial Intelligence Laboratory at Rutgers University, explains that because AI systems are designed to generalize, they naturally gravitate toward familiar themes in their training data 1

1

. The convergence may partly reflect how visual model datasets are typically curated to be visually appealing, broadly acceptable, and free of offensive material—inadvertently creating Eurocentric visuals that flatten creative diversity.

When researchers extended the experiment to 1,000 iterations, most image sequences remained stuck once they reached one of the 12 dominant motifs. In rare cases, a trajectory abruptly jumped—moving from a snow-covered house to cows in a field and then to a quaint town—but such variations remained exceptions 1

1

. Across 1,000 different iterations of the visual telephone experiment, the pattern held firm 2

2

.

Human Oversight Becomes Critical

Jeba Rezwana, a human-AI co-creativity researcher at Towson University, says the study provides evidence that unsupervised AI systems amplify existing biases, underscoring the need to keep human oversight in the creative loop 1

1

. The findings arrive at a moment when AI models are increasingly deployed as independent agents that autonomously generate, critique, and revise multimedia content. Even a simple question to ChatGPT can trigger a chain reaction as one AI system hands off queries to others.

Caterina Moruzzi, a philosopher studying creativity and AI at the Edinburgh College of Art, notes a key distinction: while human cultures develop corrective countercultures that push back against creative homogenization, AI convergence is "driven by reinforcement without critique" 1

1

. This lack of cultural pushback means AI systems lack true creative originality, instead recycling familiar patterns without the taste or judgment that characterizes human-made art 3

3

.

Implications for Creative Industries

The study's findings matter for anyone relying on AI for creative work. As more systems autonomously generate and judge other AI creative output, the resulting "bland soup of cliches" threatens to flatten the visual landscape 1

1

. The research quantifies what many have sensed intuitively—that uncanny feeling when encountering AI-generated imagery that something artificial is afoot 3

3

.

What remains unclear is whether certain visual endpoints prove more stable than others, or how frequently systems might escape these gravitational wells. "Does everybody end up in Paris or something? We don't know," Hintze admits 1

1

. The discovery that AI creativity can be condensed to just 12 motifs underscores how important human creativity and artistic judgment remain in an era of increasingly autonomous generative systems.