According to science.org, researchers from Dalarna University published a study today in the journal Patterns where they had AI models play 100 rounds of “visual telephone.” They used deliberately wild and distinct text prompts to start, feeding them into the image generator Stable Diffusion XL. The resulting image was described by another AI, the Large Language and Vision Assistant, and that new description was fed back to create another image, repeating the cycle. Across all prompts, the AI outputs consistently converged on the same 12 generic, often Eurocentric visual motifs, like pastoral landscapes or Gothic cathedrals. The researchers warn this “visual elevator music” could flatten creative diversity as AI systems are increasingly used to autonomously generate and judge creative work.
The Bland Soup Problem
Here’s the thing: this isn’t just a quirky lab result. It’s a direct peek into what happens when you take humans out of the creative loop. The AIs weren’t trying to be boring. They were just optimizing for what their training data and algorithms reward: the most statistically probable, easily describable, and broadly “acceptable” image. It’s like two people who only know how to say “nice” and “good” trying to have a deep conversation about art. They’ll quickly default to the safest, most common ground.
And that common ground, as the study shows, is weirdly specific yet utterly generic. Rainy nights in Paris. Pompous sitting rooms with chandeliers. It’s the visual equivalent of stock photo hell. The lead researcher, Arend Hintze, perfectly called it the “meaningless, happy nonsense” you find in IKEA picture frames. That’s not an accident. The datasets these models are trained on are curated to be inoffensive and visually appealing to a broad, often Western, audience. So the AI isn’t exploring; it’s regressing to the mean. Hard.
Why This Is Scary for Autonomous AI
Now, the big concern isn’t that one AI can make a cliché image. We’ve seen that. It’s that we’re building systems where AIs are set up to judge and refine each other’s work autonomously. Think of all those AI “agents” we keep hearing about. One AI generates a marketing image, another AI critiques it, a third revises it. This study shows that chain reaction doesn’t lead to innovation. It leads to a bland soup where every idea eventually drifts to one of twelve boring endpoints.
As philosopher Caterina Moruzzi pointed out, human cultures have countercultures. There’s pushback against homogenization. But in this AI loop, “convergence is driven by reinforcement without critique.” There’s no reward for being weird or challenging. The system inherently reinforces sameness. So what happens when we let these systems loose to generate content for websites, ads, or even preliminary design work without a human in the driver’s seat? We might just get a world that looks like a very boring, very beige stock photo gallery.
Is This Even a Problem to Solve?
This is where it gets philosophical. Christian Guckelsberger at Aalto University hopes we don’t just see this as an “engineering challenge” to fix. I think he’s onto something. The study forces us to ask: what is creativity for in an AI context? Is the goal to make a machine that can produce endless, mildly pleasing variations of a cathedral? Or is creativity fundamentally a human act of meaning-making and self-realization that we’re outsourcing at our own peril?
Basically, the AI found a local optimum—a creative dead end that’s comfortable and easy to reproduce. The fact that in a 1000-iteration test one sequence suddenly jumped from a snowy house to cows shows there might be fleeting escapes, but the pull to the familiar is overwhelming. The researchers admit they don’t know if everyone ends up in Paris, but the odds aren’t good. So maybe the question isn’t how to make AI more creatively divergent. Maybe it’s why we’re so eager to automate a process that, at its best, is gloriously human and unpredictable.
