We’ve been sold a grand narrative about artificial intelligence, haven’t we? The story goes that these new models are not just calculators, but creators. That a new da Vinci or Picasso is gestating in the silicon, ready to unleash a renaissance of digital art. It’s a compelling tale, but like many stories coming out of Silicon Valley, it might be more fiction than fact. The conversation around generative AI creativity has been dominated by wonder, but it’s time for a dose of pointed reality. Is this the creative revolution we were promised, or are we just witnessing a very sophisticated form of digital karaoke?
A recent study, brilliantly reported by Gizmodo, has peeled back the curtain, and what it reveals is less about an emerging artist and more about a predictable machine. Researchers set up a simple but ingenious experiment, a game of ‘visual telephone’. It’s an updated version of the party game where a message gets distorted as it’s passed along. In this case, they gave an image generator, Stable Diffusion XL, a text prompt. The AI created an image. Then, a vision model, LLaVA, described that new image in words. That description was fed back into Stable Diffusion to create the next image. They repeated this cycle 100 times and ran the entire experiment a thousand times over.
What do you suppose happened? Did it spiral into infinite abstract wonders or bizarre, never-before-seen worlds? Not quite.
The Great Creative Convergence
The creative tool evolution from basic photo filters to complex models like Stable Diffusion has been genuinely staggering. These tools can produce stunning visuals in seconds, acting as a powerful assistant for artists and designers. Yet, this fascinating study exposes a fundamental weakness in this evolution.
Regardless of the starting prompt—whether it was “an astronaut riding a unicorn” or “a serene Japanese garden”—the AI’s output consistently and inexorably converged. Across 1,000 separate chains, the images all began collapsing into one of just 12 dominant visual motifs. It’s a shockingly small number. By the 100th turn in the game, the AI had settled into its comfort zone, producing variations of the same dozen themes.
These themes included things like:
– Maritime scenes with lighthouses
– Opulent, formal interior rooms
– Moody urban settings at night
– A doe in a sun-dappled forest
Pushing Against Artistic Expression Boundaries
This convergence exposes the very real artistic expression boundaries of current AI. A human playing this game would do the opposite. Our individual biases, memories, and quirky interpretations would cause the images to diverge wildly. An astronaut might become a deep-sea diver, a unicorn might morph into a winged horse, and then a mythological griffin. Our creativity is a process of divergence.
AI’s ‘creativity’, it turns out, is a process of convergence. It defaults to the safest, most statistically probable representation of a concept based on its training data. Think of it like a river flowing downhill; it will always follow the path of least resistance to the largest body of water. For AI, that path leads directly to the most over-represented images on the internet. The researchers from the University of Wyoming and the University of Padua perhaps put it best, describing the output as “visual elevator music”—aesthetically pleasing on a superficial level, but ultimately bland, generic, and forgettable. It’s the sort of inoffensive art you’d find hanging in a chain hotel room.
The Unsolved Puzzle of Algorithmic Originality
So, why does this happen? The answer lies in the algorithmic originality challenges that are baked into the very architecture of these models. An AI image generator doesn’t understand a lighthouse in the way a human does. It doesn’t associate it with salty air, the lonely life of a keeper, or the symbolism of guidance in a storm.
Instead, it has processed millions of images tagged with the word “lighthouse” and calculated a mathematical average—a sort of ‘platonic ideal’ of a lighthouse. When pushed through iterative cycles, it regresses to that mean. It’s not creating; it’s pattern-matching at an epic scale. It’s like a musician who has perfectly memorised every pop song ever written but cannot compose a single original melody. They can endlessly remix and rearrange what already exists, but they can’t create the next truly new sound. This is the core of the problem: generative AI creativity is statistical, not conceptual.
A New Canvas: The Human-AI Collaboration Potential
Does this mean these tools are useless? Absolutely not. To dismiss them would be as shortsighted as dismissing the invention of the camera because it “only captures reality”. This study doesn’t spell the end of AI in art; it clarifies its role.
The immense human-AI collaboration potential lies in recognising AI not as a creative replacement, but as a powerful, sometimes unpredictable, creative partner. Its limitations are, in fact, an invitation for human intervention. The AI can generate a thousand ideas in a minute—990 of which might be derivative ‘elevator music’—but a human with taste, vision, and a story to tell can spot the ten sparks of potential and guide them into something truly original. The future isn’t the AI artist; it’s the human director working with a tireless AI image-maker.
This technology is a tool, and like any tool, its value is determined by the skill of the person wielding it. A paintbrush can be used to whitewash a fence or to create the Mona Lisa. The same applies here. The AI can provide the raw visual material, but the curation, the narrative, and the emotional resonance must come from us. As this research shows, leaving the AI to its own devices results in a bland echo chamber.
The hype cycle is fun, but real progress happens when we get honest about a technology’s capabilities and its limits. We are still in the very early days of this creative tool evolution. The challenge now is to move beyond the novelty of “look what an AI can make” and start asking, “what can we make with this AI?”. The most interesting art won’t come from a prompt; it will come from a process where a human is directing, editing, and challenging the AI’s tendency to settle for the average.
So, the next time you see a stunning AI-generated image, the question shouldn’t just be “what was the prompt?”. The more important question is, where is the human fingerprint? Where is the taste, the insight, the story? Because that’s where true creativity will always be found.
What do you think? Is this study a fatal blow to the idea of AI artists, or just a temporary hurdle? Let me know your thoughts below.
