Generative AI Is Flattening Culture — Study Warns of “Visual Elevator Music”

Generative AI systems are trained on vast collections of human-created web content. New research warns that when those systems begin to rely on AI-generated material, a dangerous feedback loop can form that steadily erodes variety and richness in creative outputs.

In a study published this month in the journal Patterns, an international team connected a text-to-image generator to an image-to-text model and instructed them to iterate repeatedly. Over many cycles the paired systems converged on a narrow set of bland, generic visuals the authors call "visual elevator music." Crucially, this drift toward sameness occurred without any retraining or addition of new data — the collapse emerged from repeated autonomous use.

What the experiment found

The researchers observed that autonomous feedback loops naturally gravitate toward common attractors: statistically average, describable, and easy-to-reproduce outputs. As models iteratively consume and regenerate machine-produced content, details fade and artifacts accumulate, producing images and captions that are increasingly generic and sometimes mangled.

"This finding reveals that, even without additional training, autonomous AI feedback loops naturally drift toward common attractors," the authors wrote. "Human-AI collaboration, rather than fully autonomous creation, may be essential to preserve variety and surprise in the increasingly machine-generated creative landscape."

Broader cultural concerns

Rutgers computer science professor Ahmed Elgammal, writing about the paper for The Conversation, argues the study adds to mounting evidence that generative AI is already contributing to a form of cultural stagnation. He warns that algorithms—by promoting familiar and easily describable content—may push AI outputs to the top of online feeds, displacing diverse human-created work.

"The convergence to a set of bland, stock images happened without retraining. No new data was added. Nothing was learned. The collapse emerged purely from repeated use," Elgammal wrote.

The implications extend beyond images: if future training corpora contain large amounts of AI-generated content, new models could inherit and amplify the same homogenized tendencies. That raises urgent questions for creative fields from photography and illustration to advertising and film.

Possible responses

The paper and commentators recommend several interventions to preserve cultural variety:

• Prioritize human-AI collaboration so humans remain central to creative decisions.
• Design models and incentives that resist convergence toward statistically average outputs, encouraging novelty and risk-taking.
• Monitor and label AI-generated content to reduce the risk of training-on-synthetic-data feedback loops.
• Support diverse human-created datasets and curation practices to counterbalance mass-produced AI content.

Without such measures, researchers warn, generative systems may continue to drift toward mediocre, uninspired results — a slow but steady flattening of cultural richness. The study is a timely reminder that technical progress must be paired with thoughtful design and policy to keep AI from narrowing the very creative landscapes it promises to expand.