According to Futurism, a new study published this month in the journal Patterns by an international team of researchers found that text-to-image generators, when linked with image-to-text systems and set to iterate repeatedly, collapse into producing “very generic-looking images” dubbed “visual elevator music.” The research, highlighted by Rutgers University professor Ahmed Elgammal in The Conversation, shows this convergence happens without any new training data, leading to what he calls “cultural stagnation.” The study argues that AI systems are already operating this way by default, with the collapse emerging purely from repeated use. This is happening as a tidal wave of AI-generated content floods the internet, and algorithms begin to favor this homogenized output. The researchers conclude that human-AI collaboration, not fully autonomous creation, is essential to preserve variety.
The Feedback Loop Problem
Here’s the thing that’s so unsettling about this research. It’s not a prediction about some far-off future. The study shows the degradation happens right now, just by having an AI talk to itself. You set up a text-to-image model and an image-to-text model in a loop. The first one makes a picture from a prompt, the second one writes a new prompt based on that picture, and so on. After a few rounds, you’re left with mush. Bland, stock-image nonsense. The technical term is they drift toward “common attractors.” I think a simpler term is creative death spiral.
And the scary part? “No new data was added. Nothing was learned.” The AI isn’t getting smarter or more creative. It’s getting dumber and more repetitive, like a game of telephone played by a machine that only understands averages. This is the core of the cultural stagnation argument. If future AI models are trained on this ocean of AI-generated slop—which is already happening—they’ll start their lives already biased toward the generic. It’s a vicious cycle with no obvious off-ramp.
Why This Is a Business Problem Too
So, you might think, “Big deal, artists should just not use AI.” But that misses the scale of the issue. Look at the business model for most generative AI companies. It’s built on volume, speed, and cost-cutting. They’re selling the promise of replacing human creative labor. But what are you replacing it with? According to this research, you’re replacing it with a system that inherently trends toward the mediocre.
That’s a terrible long-term product strategy. If every marketing image, blog post, and social media ad starts to look and sound the same—a sort of beige paste of content—then the tool loses its value. The initial cost savings get wiped out by a total lack of competitive differentiation. Who wants a tool that makes everything look like everything else? The beneficiaries in the short term might be companies selling the AI tools, but the long-term losers are everyone: the platforms flooded with boring content, the businesses that can’t stand out, and obviously, the human creators pushed out of the market.
Is There a Way Out?
The researchers, and Elgammal, point to one potential solution: human-in-the-loop systems. Basically, AI as a collaborator, not a replacement. The models need to be designed to resist their own urge to converge on the average. They need incentives to be weird, to deviate, to surprise us. That’s a much harder technical and philosophical problem than just scaling up parameter counts.
But let’s be skeptical for a second. The entire economic incentive right now is for autonomy and scale. “Fully automated” is the buzzword. Building systems that require human judgment to stay creative seems to run counter to the sales pitch. Can the industry pivot from “replace humans” to “augment humans” in time to avoid poisoning its own data well? I’m not sure. The essay in The Conversation ends on a hopeful note about design interventions, but the current trajectory feels pretty firmly set.
The Mediocrity Machine
This isn’t just about art. It’s about information, communication, and culture itself. Elgammal warns the risk is “that AI-mediated culture is already being filtered in ways that favor the familiar, the describable and the conventional.” Think about that. Our collective culture—our stories, our visual language—could be shaped by a machine’s bias toward the statistical middle.
We’re building a perfect mediocrity machine. And we’re feeding it our entire past to produce what it thinks is our future. The study, which you can read more about in the original paper, makes it clear this isn’t speculation. It’s already happening. The question is whether we care enough about variety and surprise to try and stop it.
