A bold claim now backed by mathematics: large language models-the engines behind generative AI systems such as ChatGPT-are structurally unable to achieve expert-level creativity. The finding, from David H. Cropley, Professor of Engineering Innovation at the University of South Australia, has been published in the Journal of Creative Behavior and reframes the debate over whether AI can rival human ingenuity. In his analysis, these systems reach a hard ceiling at a creativity score of 0.25 on a scale from zero to one-a level corresponding to the boundary between “little-c” amateur creativity and “Pro-c” professional competence.
Cropley’s approach was based on the standard definition of creativity: A product must be both effective-useful, appropriate, and fit for purpose-and original-novel, unusual, and surprising. In human high-level creativity, these qualities co-occur; a great invention is both singular and flawlessly executed. But in the probabilistic mechanics of large language models, the qualities are locked in a trade-off. The “next-token prediction” process by which the model calculates the most probable word or token to follow in a sequence inherently ties effectiveness to statistical likelihood. Selecting a highly probable token ensures coherence but erodes novelty; selecting a rare token boosts novelty but often undermines sense and utility.
This trade-off is not only empirical but also mathematically expressible. Cropley modeled creativity as a product of effectiveness and novelty, each inversely related in a closed probabilistic system. The result is a maximum achievable score of 0.25, achieved only when both variables sit at moderate levels. This means that LLMs cannot simultaneously maximize originality and effectiveness, a feat human experts achieve as a matter of course. In practice, this cap aligns with the empirical data showing AI-generated stories and solutions rank in the 40th–50th percentile compared to human outputs.
The mechanics driving this ceiling borrow from information theory, in which novelty can be quantified as a deviation from expected statistical patterns. Trained on immense corpora of human text, LLMs operate within the distribution of their training data. Even when their outputs seem surprising to casual observers, they remain recombinations of familiar structures. That is why highly creative professionals can detect formulaic tendencies patterns, tropes, and syntactic rhythms that give away the model’s statistical roots so quickly.
Cropley’s work further emphasizes how decoding strategies affect AI creativity. Most LLM deployments are based on greedy decoding or simple sampling; these methods favor high-probability tokens, thus leaning toward effectiveness at the expense of originality. Advanced strategies, such as nucleus sampling or temperature scaling, introduce more randomness, thus nudging novelty upward. Even with these adjustments, however, the underlying trade-off persists, and the ceiling remains. Architectural changes are needed that break dependence on past statistical patterns without which these tweaks can only shift the balance point within the same constrained space.
Emerging research into alternative architectures might address this bottleneck. Some experimental systems attempt to integrate generative processes that are not strictly tethered to token probability distributions, potentially allowing for outputs that escape the statistical gravity of the training data. Others explore hybrid models that mix symbolic reasoning with neural generation, hoping to inject structured novelty without sacrificing coherence. Yet the conclusion from Cropley is clear: under current design principles, no matter the decoding method, the mathematical limit holds.
The implications go beyond academic curiosity. Industries that might be tempted to automate creative labor – advertising, entertainment, product design will risk homogenizing their output if they rely too heavily on the LLMs. Since roughly 60% of people score below average on creativity tests, many will find AI output impressive. For sectors in which transformative originality drives value, though, this ceiling signals danger: over-reliance on AI will lead to formulaic, repetitive work – eroding competitive differentiation.
As Cropley puts it, “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.” For AI to ascend to expert levels of creativity, it would have to be founded on fundamentally new architectures capable of creating ideas disconnected from prior statistical patterns-a sea change in computer science that is still beyond the horizon.
