The Creative Gap: AI Can Generate Options, but Humans Still Change the Rules

TL;DR for operators

Generative AI can produce a pile of plausible options before a human team has finished developing one. The harder question begins afterward: which option is actually interesting, which is merely competent, and when does an unexpected result deserve to change the direction of the work?

Ivan Magrin-Chagnolleau’s Can an AI System Be Creative? A Critical Perspective from Art and Engineering¹ makes that gap visible in a haiku exercise. Across six batches, the AI produced 60 poems that followed the requested form, yet repeatedly converged on probable structures and was weak at distinguishing its strongest outputs. The paper’s argument is therefore not simply that AI lacks novelty. It separates rapid generation from the harder capabilities of judging creative value and recognizing when an accident is significant enough to revise the original objective.

That distinction leads to a more useful view than either “AI is creative” or “AI only copies.” Contemporary systems can recombine existing material, explore many possibilities within an established problem space, and sometimes surface results that redirect human intention. The stronger boundary appears when creativity requires changing the rules that define the space itself: the paper argues that today’s systems lack autonomous creative-value judgment and the subject position needed to recognize an unexpected event as meaningful and choose to reorganize the work around it.

For design, media, marketing, and R&D teams, the operational implication is a division of labor: use AI to expand and perturb the option space through ideation, recombination, prototyping, and execution support, while keeping domain experts responsible for selection, redirection, and decisions about when the brief itself should change. That critical distance also matters because repeated statistically probable suggestions may pull choices toward convention rather than away from it.

The evidence boundary is substantial. The paper offers a conceptual and phenomenological argument supported by four heterogeneous illustrative cases, not a controlled multi-model creativity benchmark. Its workflow implications are therefore propositions worth designing around and testing, not measured estimates of how AI use affects originality.

Sixty acceptable haikus expose the harder part of creativity

A familiar generative-AI workflow begins with abundance. Ask for alternatives and the system can return ten, twenty, or fifty reasonably formed possibilities faster than a human team could develop them manually.

The paper’s haiku exercise makes the resulting ambiguity unusually clear. Across six batches, the AI generated 60 poems that respected the requested form. Yet the outputs repeatedly converged on probable structures, and the system was weak at distinguishing its strongest work.

That combination matters more than either observation alone. Formal competence demonstrates generative capability. Repetition and weak self-ranking expose a separate problem: producing alternatives does not establish that the producer can determine which alternative deserves creative attention.

The literary rewriting case reinforces the distinction. Seven variants of one autobiographical paragraph demonstrated strong stylistic imitation, but the author’s assessment found pastiche, caricature of surface features, loss of individual authorial voice, and no internally grounded hierarchy among the results.

These cases are illustrative evidence for the paper’s conceptual thesis, not controlled comparisons, robustness tests, or statistical demonstrations of model behavior. Their value is diagnostic. They separate fluency from judgment.

Creativity has different depths

Once that separation is established, the paper introduces Margaret Boden’s framework.

At the first level, combinatorial creativity creates new arrangements from existing material. Contemporary AI is well suited to this. Large training corpora and fast generation make systems effective at mixing stylistic elements, concepts, structures, and representations.

A deeper capability is exploratory creativity: searching through possibilities permitted by an existing conceptual space. AI can also do substantial work here. It can rapidly traverse variants that a person might never have time to enumerate.

The paper places the stronger boundary at transformational creativity. Some creative advances do not simply locate a better point inside the existing space; they alter the assumptions or rules that determine what belongs in that space at all. On the paper’s account, contemporary AI cannot perform this transformation in the strongest sense.

This produces a more useful classification than either “AI is creative” or “AI only copies.” A system can materially expand the range of available options without possessing every capability involved in deciding what the creative problem should become.

The constraint is generation, evaluation, and redirection together

The paper proposes several mechanisms behind that boundary.

First, contemporary generative systems depend on previously observed data and probabilistic generation. The author’s argument is that this creates a gravitational tendency toward familiar structures: outputs can be novel in combination while still being statistically pulled toward convention.

Second, creativity requires value, not novelty alone. A model can reproduce learned patterns of human preference or score outputs against externally supplied criteria. The stronger claim here is narrower: this is not equivalent to autonomously determining that a particular result is aesthetically, intellectually, emotionally, or practically worth pursuing.

Third comes the paper’s most distinctive contribution. An unexpected event becomes creatively consequential only when someone notices that it matters and allows it to redirect intention. Random variation by itself is therefore insufficient.

This is the role assigned to serendipity. The critical capability is not simply producing something unforeseen. It is recognizing the unforeseen result as significant, abandoning or revising part of the previous plan, and following the new possibility.

Under this account, transformational creativity requires a coupling between surprise, evaluation, and agency. The paper locates that coupling on the human side of today’s human-AI interaction.

AI can still change what the human creates

The conclusion is therefore not that AI contributes little to creative work.

A previously published computer-assisted music-composition case shows an algorithm functioning as a productive constraint. A luminosity curve extracted from a 1m30s film excerpt imposed material that displaced the human composer from habitual choices. The system did not need to be credited with composing the resulting work for the constraint to change the human creative trajectory.

The image-generation cases show a related dynamic. Iterative Ideogram outputs sometimes revealed possibilities that altered what the human author subsequently wanted. The paper describes this as creative co-evolution: intention changes in response to generated possibilities rather than remaining fixed from the first prompt.

These cases illustrate a stronger role for AI than simple automation. The system can perturb the search space, expose alternatives, and execute possibilities cheaply enough that the human encounters routes that might otherwise have remained invisible.

But the decisive operation remains selection. Someone must recognize that an unexpected output is not merely different but worth following.

Creative teams should design for option expansion, not delegated taste

For organizations, Cognaptus draws a workflow implication rather than a claim about machine consciousness.

This suggests that the quality of an AI-assisted workflow should not be measured only by how much content the model produces or how polished the first result appears.

A stronger process preserves expert checkpoints at moments of evaluation and redirection. Prompt literacy then extends beyond writing better instructions. It includes diagnosing how the model interpreted an intention, comparing alternatives, noticing divergence, and recognizing when an unintended result is more promising than the original request.

That places substantial value on domain expertise. Someone unable to distinguish a conventional answer from an unusually good one gains less from receiving a larger pile of plausible options.

The averaging risk is plausible, but not quantified here

The paper also warns that repeated exposure to statistically probable suggestions can pull creative work toward convention, cliché, and polished average solutions.

The mechanism is coherent with the haiku and literary examples: systems capable of producing competent patterns at scale may repeatedly surface solutions near established regularities. If humans increasingly choose among those suggestions, the model could influence not only execution but the distribution of options considered.

For operators, that is a reason to preserve critical distance and sometimes use AI deliberately as a displacement mechanism rather than as a finishing machine.

It is not, however, evidence that AI use has been shown to homogenize creative industries. The four cases span different media, periods, and systems; several judgments depend on the author’s aesthetic assessment; there is no standardized creativity metric, blinded evaluation, inferential analysis, or multi-model comparison. One literary experiment also lacks a recorded ChatGPT version.

The paper’s strongest claims concern contemporary systems as characterized in its conceptual analysis. They do not empirically settle what systems with different forms of learning, embodiment, autonomy, or self-modification could eventually do.

The useful boundary is who can change direction

The paper’s contribution is less a verdict on whether an AI “is creative” than a decomposition of the work hidden inside that word.

Current systems can contribute substantial combinatorial power. They can explore possibilities quickly. They can impose constraints, execute unfamiliar variations, and present outputs that cause a human collaborator to reconsider an initial intention.

What the paper reserves for the human is the consequential transition from unexpected output to changed direction: recognizing value, deciding that an accident matters, revising the objective, and accepting responsibility for the resulting choice.

For organizations adopting generative AI in creative work, that boundary leads to a concrete design principle. Expand machine involvement where more alternatives, cheaper experimentation, or productive displacement improves the search process. Preserve human authority where the task changes from generating possibilities to deciding which possibilities should redefine the work.

Cognaptus: Automate the Present, Incubate the Future.