AI image generators may boast huge training data and promise endless creativity, but it turns out there are limits to their imagination. According to new research, when pressured to constantly reinterpret scenes, these models quickly abandon originality and settle into a handful of predictable, almost clichéd styles. The study, published in the journal Patterns, shows that no matter how detailed the initial prompt is, AI image generators like Stable Diffusion XL and LLAVA revolve around the same dozen artistic motifs. Think lighthouses, city skylines at night, rustic architecture, and general indoor scenes, the kind you might find hanging in a hotel lobby rather than an art gallery.

Researchers have described this effect as “visual elevator music”, images that feel polished yet hollow, completely ineffective but ultimately forgettable. To test the AI’s creativity, researchers designed a game they call Visual Telephone. The setup was simple, at least in theory. One model, the Stable Diffusion XL, was given a poetic allusion, such as, “When I was particularly sitting 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 generated an image from that text, which was handed off to another model, LLAVA, to describe. The detail produced by LLAVA was fed back to Stable Diffusion, which used it to create a new image. This process was repeated 100 times in a digital echo chamber from image to description to new image.

Like the childhood game of telephone, where a message gets distorted as it passes from one person to another, the original concept quickly disappeared. By the tenth or twentieth round, any resemblance to the initial image had faded.

However, what shocked the researchers was not just the distortion, it was the convergence. In 1,000 different iterations of the experiment, the models consistently gravitated toward 12 recurring styles. No matter how poetic, strange, or abstract the initial prompt, the end result always follows one of the same motifs.

The researchers observed that stylistic change generally occurred gradually, with the images losing their distinctiveness one by one. Sometimes, change came suddenly, a sudden decline in something. But the destination was always the same: bright, generic views with a mildly familiar feel.

Even when the team transformed the models using different versions of both the image generator and description tools, the trend persisted. Extending the experiment to 1,000 turns only strengthened the pattern. Around the 100th image, the sequence will settle into a particular motif, and subsequent iterations will produce small variations on that theme.

“In 1,000 different iterations of the telephone game, the researchers found that most image sequences would eventually fall into one of 12 major motifs,” the study said. In other words, the machines had a comfort zone and they liked to live in it.

So what does this say about artificial creativity? The researchers argue that, while humans often offer unexpected explanations, AI irons out the irregularities. “In the human game of telephone, you will end up with enormous variation because each message is delivered and heard differently,” the paper reported. “AI has the opposite problem. No matter how strange the original prompt is, it will always default to a narrow selection of styles.”

AI, of course, learns from human-generated data, which can also contribute to this creative loop. If most photos online fall into similar categories, sunsets, streets, interiors, it’s no surprise that AI finds itself stuck repeating familiar visual tunes.