Why Machines Can’t Match Our Wildest Ideas

Creativity has never been a numbers game, and a new Australian analysis offers a stark reminder of just how far generative AI sits from human imagination. In a study grounded in mathematics, researchers show that today’s large language models hit a ceiling long before they reach the ingenuity of society’s most inventive thinkers.

The paper, led by University of South Australia creativity expert David Cropley and published in the Journal of Creative Behaviour, uses the internal mechanics of large language models to calculate their maximum creative potential. Cropley finds that the architecture behind systems like ChatGPT mathematically caps their creativity at 0.25 on a standard scale of 0 to 1, placing them only at the threshold between amateur and professional performance. His team conducted a formal data and statistical analysis and concludes that under current design principles, LLMs can mimic creativity but cannot reach expert levels.

Cropley argues that misunderstanding fuels much of the hype surrounding generative AI. People see stories, poems, or images pop out of an algorithm and assume that creativity has taken root. But he points out that these systems remix statistical patterns rather than originating unfamiliar or effective ideas. The distinction matters, he says, because originality sits at the heart of every definition of creativity used in psychological research. By tracing how LLMs assemble predictable outputs from past data, his analysis reframes the public debate over what machines are truly capable of.

Patterns, Not Possibilities

“Many people think that because ChatGPT can generate stories, poems or images, that it must be creative. But generating something is not the same as being creative. LLMs are trained on a vast amount of existing content. They respond to prompts based on what they have learned, producing outputs that are expected and unsurprising.”

The study emphasizes that LLMs operate within a symmetrical mathematical space, producing an average of what they have already seen rather than leaping toward the unexpected. This has consequences not only for how society evaluates AI, but also for industries increasingly tempted to automate creative labor. If the underlying architecture cannot escape its training distribution, Cropley warns, then organizations that depend too heavily on generative systems risk homogenizing their output. For fields that rely on novelty, from entertainment to product design, that could mean a slow drift toward uniformity.

Cropley also notes a psychological wrinkle. Because roughly 60 percent of people fall below average on creativity tests, it is statistically inevitable that many will perceive machine output as impressive or even innovative. Highly creative individuals, he says, recognize the limitations more readily, spotting the telltale patterns and familiar forms that algorithms cannot break free from.

The Human Edge

“A skilled writer, artist or designer can occasionally produce something truly original and effective. An LLM never will. It will always produce something average, and if industries rely too heavily on it, they will end up with formulaic, repetitive work.”

Cropley sees the findings as good news. Far from replacing playwrights, poets, or painters, generative AI remains a tool that augments rather than rivals extraordinary human talent. For machines to reach expert-level creativity, he writes, they would require entirely new architectures capable of breaking free from past statistical patterns. That threshold remains out of reach for current systems.

For now, the takeaway is simple. Human imagination, with all its risk taking and occasional brilliance, continues to set a standard that machines cannot meet. And according to Cropley, that creative spark is something the world needs now more than ever.

Journal of Creative Behaviour: 10.1002/jocb.70077