Stable Diffusion can whip up a photorealistic dragon riding a skateboard through a cyberpunk Tokyo in about eight seconds. Your art school friend takes three weeks to finish a still life of pears. Yet somehow, when researchers put both to the test, the flesh-and-blood creator still came out ahead on the creativity scoreboard.

A new study published in Advanced Science pitted humans against AI in a visual creativity showdown, and the results spell out a clear pecking order: trained visual artists crush it, non-artists hold their own, and AI image generators trail behind - especially when left to their own devices.

The Experiment: Artists, Amateurs, and Algorithms Walk Into a Lab

Researchers gathered 27 professional visual artists and 26 regular folks (the kind who haven't touched a paintbrush since kindergarten finger painting). Both groups completed image generation tasks. Then they had Stable Diffusion do the same thing under two conditions: "Human-Inspired" (where the AI got detailed guidance from humans) and "Self-Guided" (where the AI was basically told "go wild, buddy").

Then came the judges. A panel of 255 human raters scored all the resulting images on liking, vividness, originality, aesthetics, and curiosity. And because we apparently live in an era where AIs judge other AIs, GPT-4o also weighed in as a rater.

The creativity gradient that emerged was stark: Visual Artists > Non-Artists ≥ Human-Inspired GenAI > Self-Guided GenAI. Translation: without a human holding its hand, the AI's creative output tanked.

Why Can AI Write Poetry But Struggle With Pictures?

Here's where things get weird. Large language models have been shown to match or exceed average human performance on divergent thinking tasks - the psychological tests that measure how many different ideas you can generate from a single prompt. GPT-4 can riff on word associations all day long and score above the human average.

But visual creativity operates on different rules. It depends on what the researchers call "perceptual nuance and contextual sensitivity" - fancy terms for knowing that putting a tiny red balloon in the corner of a grey industrial scene creates emotional tension, or understanding why a particular brush stroke feels melancholic. These skills didn't transfer from language models to image generators like everyone hoped.

Divergent thinking, first defined by psychologist J.P. Guilford in 1956, involves generating multiple diverse solutions in a free-flowing, non-linear way. AI can do this with words. With images? Not so much, unless humans provide the creative scaffolding.

The AI Judge Had Some Explaining to Do

Perhaps the most eyebrow-raising finding: GPT-4o as a rater behaved... questionably. Without specific guidance, it inflated scores for AI-generated images and showed "reduced discrimination between image categories." It basically gave participation trophies to its silicon cousins.

Only when researchers trained GPT-4o with examples of how humans actually rate creativity did it start making judgments that aligned with human evaluators. Left unguided, the AI rater essentially flattered other AIs while underselling human work. Research has shown that LLMs exhibit evaluation biases, sometimes favoring outputs that match patterns in their training data rather than genuinely assessing quality.

What This Means for Your Creative Workflow

The takeaway isn't "AI bad, humans good" - it's more nuanced than that. Human-guided AI output came close to matching non-artist humans. The bottleneck isn't the technology itself; it's the creative direction.

Think of Stable Diffusion as an extremely talented but directionless intern. Hand it a detailed brief, mood boards, and specific references, and it produces impressive work. Tell it to "make something cool" and you get the visual equivalent of clip art with extra steps.

For anyone working with AI image tools, this validates what power users already know: prompt engineering isn't just tweaking keywords. It's transferring your creative vision into a format the model can execute. The creativity still has to come from somewhere, and right now, that somewhere is you.

The Creative Hierarchy Holds - For Now

This research arrives at an interesting moment. A separate large-scale study testing over 100,000 people found that while AI can beat the average human on certain creativity tests, the top 10% of creative humans remain untouchable - particularly for complex work like poetry and storytelling.

The pattern holds across domains: AI performs impressively at the median but falls apart at the extremes. The most creative humans still produce work that machines can't match, while unguided AI settles into a kind of competent mediocrity.

For visual artists worried about obsolescence, this research offers some comfort. Your perceptual instincts, contextual awareness, and ability to make creative choices that resonate emotionally - those remain distinctly human capacities that don't emerge just because you train a model on billions of images.

The machines paint pretty pictures. But the art still needs an artist.

References:

• Rondini, S., et al. (2025). Stable Diffusion Models Reveal a Persisting Human - AI Gap in Visual Creativity. Advanced Science. DOI: 10.1002/advs.202524142 | arXiv:2511.16814
• Hubert, K.F., et al. (2025). A large-scale comparison of divergent creativity in humans and large language models. Nature Human Behaviour. Link
• Zhou, E. & Lee, D. (2024). Generative artificial intelligence, human creativity, and art. PNAS Nexus. PMID: 38444602

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.