Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles

The advancement of generative AI has brought with it arguments over just how much “creativity” a machine can have, but a group of researchers recently published a study that suggests generative AI image makers tend to default to just 12 general motifs or topics.

In the study, which the authors describe as “sobering for computational creativity,” two AI systems were set up to generate 100 images without human input. Despite repeating that test to generate thousands of images, the researchers found that the generations all tended to fall into one of 12 common motifs, such as a bridge, a lonely tree, an action photo, or a lighthouse scene.

The researchers described the resulting collection of images as generic-looking “visual elevator music.” The study comes from research based out of the Department of Data Analytics at Dalarna University in Sweden and the BEACON Center for the Study of Evolution in Action at Michigan State University in the US.

For the research, the group wanted to set up an AI to generate images without human input. To do that, the team gave an AI image generator a prompt to produce an image. A second AI was then asked to describe that image. The description was then fed back into the image generator to create an image based on that AI-written prompt. This went on for 100 rounds. Researchers repeated the experiment 40 times and then repeated it with four different image generators.

When the researchers looked at the results, they found that the images tended to fall into one of twelve different motifs or styles. As the researchers described it, “...the systems systematically evolved toward nearly identical semantic and visual endpoints—stormy lighthouses, urban night scenes, gothic cathedrals, and palatial interiors. Rather than exploring creative possibilities, autonomous AI loops appear to gravitate toward what could be called visual elevator music.”

“These findings are sobering for computational creativity,” the researchers wrote. “If AI systems consistently collapse toward generic outputs when operating without human intervention, this questions whether current approaches can achieve genuine machine creativity. The tendency toward ‘safe’ visual tropes suggests that maintaining creative diversity may require explicit anti-convergence mechanisms or continuous human curation.”

The consistency that the multiple image generators returned to the same 12 general topics or themes suggests “that these systems favor high-probability outputs over genuine novelty,” the researchers concluded.

The researchers warned that “widespread deployment of such systems could inadvertently homogenize visual culture.”

But, the researchers note that human culture transmission research tends to repeat common themes as well, noting themes in storytelling as well as visual arts repeated across even separate cultures and during different time periods. “The difference lies not in the presence of convergence but in the specific attractors: where humans converge on flood myths and spiral patterns shaped by embodied cognition, AI systems converge on stock photography aesthetics shaped by internet-scale training data,” the researchers explained.

Generative AI systems are trained on human-created data, the study’s authors note, wondering, “What does the convergence on common artistic motifs say about us?”

The research is available at Patterns.

You may also like

Browse the best mirrorless cameras or the best laptops for photo editing.