Artificial intelligence systems tasked with generating and refining images through autonomous iterations are increasingly falling into a pattern of producing the same repetitive, generic visuals—what researchers describe as “visual elevator music.” In a new study published in Patterns, scientists explored this phenomenon by simulating a game of visual telephone using two AI models: an image generator (Stable Diffusion XL) and an image description model (Large Language and Vision Assistant). Over 100 rounds, each image was generated from a textual prompt, described by the second model, and then used to generate a new image, creating a feedback loop. The researchers began with highly diverse and unusual prompts—such as “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”—to avoid bias from the start. Despite this effort, the AI systems rapidly converged on just 12 recurring visual themes: Gothic cathedrals, pastoral landscapes, rainy Parisian nights, luxurious sitting rooms, snow-covered houses, and other clichéd, Eurocentric scenes. One prompt about a prime minister negotiating a fragile peace deal eventually transformed into an image of a dramatic chandelier in a velvet-lined room—far removed from the original concept. The drift wasn’t due to randomness or model flaws. Even when researchers adjusted parameters or swapped in different AI models, the same 12 motifs persisted. When the experiment was extended to 1,000 rounds, most image sequences became stuck in one of these patterns, with only rare, sudden shifts—like a transition from a snow-covered house to a field of cows—breaking the cycle. The root of this homogenization may lie in the training data used to build these models. These datasets prioritize visually appealing, safe, and broadly acceptable images—often drawn from stock photo libraries—leading AI systems to favor familiar, aesthetically pleasing, but culturally narrow representations. As Ahmed Elgammal of Rutgers University notes, AI models are designed to generalize, so they naturally gravitate toward what they’ve seen most often. This trend mirrors human cultural tendencies, where certain stories and symbols recur across societies. But unlike humans, AI lacks countercultural forces or critical feedback loops to resist conformity. As Caterina Moruzzi from Edinburgh College of Art points out, AI systems are rewarded for generating stable, easily describable images—favoring sameness over novelty. The findings raise deeper concerns about the future of creativity in AI. Christian Guckelsberger of Aalto University warns against treating this as a mere technical problem to be fixed. Instead, he urges reflection on the purpose of creativity itself—its role in human meaning-making and self-expression. If AI systems are allowed to dominate creative processes without human oversight, they risk producing a global visual culture that is polished, predictable, and devoid of true originality. The real challenge, he suggests, may not be to make AI more creative, but to preserve the human need to create.