This pivotal study, published in PNAS Nexus on March 24, 2026, challenges popular assumptions about the creative versatility of LLMs. Emily Wenger, an assistant professor specializing in electrical and computer engineering at Duke, alongside Yoed Kenett, a cognitive neuroscientist affiliated with the Technion in Israel, spearheaded an experimental comparison between 22 commercial LLMs and over 100 human participants across three well-established creativity tests. Their findings reveal that, although individual models occasionally surpass single humans in creativity metrics, the collective responses generated by these models are significantly more uniform than those produced by people.

Researchers attribute this convergence to the underlying architecture and training regimes shared by commercial LLMs. These models ingest vast datasets scraped from the public internet, encompassing a broad yet overlapping corpus of human knowledge and language. The identical goal function—producing coherent, contextually appropriate text—further drives model output toward a narrow band of plausible creative avenues. Wenger observed that because all models are effectively “speaking from the same script,” their responses echo one another, limiting the diversity critical to true creative innovation.

To quantify this, the researchers employed three long-standing assessments of divergent thinking and associative richness: the Alternative Uses Test (AUT), the Divergent Association Task (DAT), and the Forward Flow (FF) test. AUT prompts examinees to conceptualize unconventional applications for everyday objects, such as envisioning a book as a makeshift doorstop or a firestarter. DAT demands the generation of ten semantically distant words, probing the breadth of associative leaps. Finally, FF entails a sequential chain of word associations stemming from a seed word, measuring the fluidity and novelty of cognitive transitions.

The choice of tests was strategic, targeting cognitive mechanisms fundamental to creativity: variability, originality, and the looseness of conceptual networks. Human responses demonstrated extensive heterogeneity in word choice and concept paths, embodying flexible, explorative thought processes. Conversely, the LLM outputs were markedly homogenized, clustering around common or high-probability associations that reflect their training “bias” toward widely observed linguistic patterns rather than unpredictable ideation.

Kenett elaborates on the implications of this phenomenon: “Human creativity thrives on variability, enabling breakthroughs through unique and often unexpected connections. If AI-generated content converges too tightly in its creative space, it risks stagnating innovation and promoting uniformity.” This homogenization is a subtle but profound limitation, potentially undermining the transformative promise of AI in augmenting human creativity. It raises a red flag concerning the overreliance on these systems in creative industries and beyond.

The study also experimented with manipulating the system prompt—the initial instructions guiding the LLMs’ behavior—to encourage more divergent and novel outputs. Yet, these manipulations only yielded marginal increases in variability, insufficient to rival human-generated diversity. This points to intrinsic constraints in the models’ design or training data that mere prompting cannot overcome. It suggests a deeper architectural or methodological innovation might be necessary to break the creative mold imposed by current training paradigms.

Given the surge in LLM adoption highlighted by a 2024 Adobe survey—reporting over 50% of Americans have engaged AI systems as creative partners in writing, coding, and ideation—this research carries urgent pragmatic warnings. The ubiquity of these tools risks cultivating a monoculture of expression, where novel concepts and linguistic creativity diminish as AI-assisted works increasingly converge stylistically and conceptually. Such an outcome challenges the narrative of AI as a catalyst for human innovation.

Wenger strongly advocates for continued human involvement in creative processes, particularly when originality and differentiation are prized. She recommends assembling diverse human teams for brainstorming to combat creative stagnation rather than relying solely on AI-generated content. The nuanced interplay of distinct human experiences and cognitive variability remains unmatched by algorithmically generated ideas, reinforcing the indispensability of human creativity in a world increasingly intertwined with AI.

This research adds a crucial dimension to ongoing discussions about the role of AI in creative disciplines, emphasizing that the promise of generative technologies must be balanced with awareness of their limitations. It calls for further interdisciplinary inquiry into the cognitive and computational factors shaping AI’s creative scope, especially as these agents become more embedded in professional and everyday workflows. Understanding and addressing AI homogeneity could pave the way for future models capable of supporting genuinely diverse, groundbreaking creativity.

In essence, while commercial large language models represent remarkable feats of engineering and linguistic fluency, their creative capacity remains intrinsically bounded by their training data, architecture, and the optimization goals set during development. As this Duke-led study compellingly reveals, the creative frontier still belongs chiefly to humans, whose unpredictable and varied cognitive styles defy the algorithmic tendencies toward uniformity. The ongoing challenge is to design AI that not only mimics human language but also fosters an ecosystem where creativity flourishes in its full, vibrant diversity.

Subject of Research: Not applicable

Article Title: Large language models are homogeneously creative.

News Publication Date: 24-Mar-2026

Web References: https://doi.org/10.1093/pnasnexus/pgag042

References: Emily Wenger and Yoed N. Kenett. “Large language models are homogeneously creative.” PNAS Nexus, 2026, 5, pgag042.

Keywords: Generative AI, Large language models, Creativity, Artificial intelligence, Divergent thinking, Cognitive variability

Historically, creativity has been dissected into various subdomains—verbal, visual, musical, and motoric—each thought to rely on different cognitive and neural pathways. However, this compartmentalized view has often struggled to explain why individuals who excel in one creative field frequently demonstrate aptitude in others. Battistello and colleagues challenge this parochial understanding by offering empirical evidence that subjective valuation serves as a core cognitive process influencing creative output regardless of the task domain. Subjective valuation here is understood as the brain’s continuous evaluation of potential ideas based on their novelty, usefulness, emotional salience, and personal relevance, fundamentally shaping which ideas are pursued and developed.

The team utilized a combination of functional magnetic resonance imaging (fMRI) and advanced computational modeling to observe participants engaged in diverse creative tasks. By measuring brain activity during ideation phases and subsequent decision-making about which concepts to expand upon, the researchers isolated common neural substrates implicated in subjective valuation decisions. Intriguingly, regions traditionally associated with reward processing and valuation—such as the ventromedial prefrontal cortex (vmPFC) and the striatum—were consistently activated across domains. These areas dynamically interacted with executive control regions like the dorsolateral prefrontal cortex, suggesting a sophisticated network that integrates evaluative and generative cognitive functions.

One of the study’s most compelling findings is the demonstration that creative cognition hinges on a continuous interplay between generation and evaluation, where subjective valuation acts as an internal feedback loop directing attention toward promising ideas. This process mirrors mechanisms seen in economic decision-making but is here repurposed to foster innovative thinking rather than maximize material gains. The authors interpret this as a domain-general cognitive algorithm specialized by experience and context, leveraging valuation to navigate immense solution spaces where objective criteria for ‘best’ ideas are often unavailable.

Beyond identifying neural correlates, the research team connected these insights to observable differences in creative performance across individuals. By correlating neural activity patterns during valuation with real-world creative achievements and psychometric creativity assessments, the study illustrates that more successful creators display heightened sensitivity in these valuation circuits. This implies that creativity, often romanticized as a spontaneous epiphany, may instead be heavily reliant on refined internal evaluative processes that selectively amplify promising concepts while discarding less viable ones.

Furthermore, the implications of these findings extend to educational and professional settings aimed at enhancing creativity. If subjective valuation is fundamental across creative domains, training programs could be designed to cultivate this evaluative capacity explicitly. Such training might include exercises that heighten metacognitive awareness of idea quality and emotional resonance, thereby empowering individuals to better navigate the ideation process. This also suggests potential benefits of neurofeedback or neuromodulation techniques targeting valuation networks to boost creative outcomes.

The study also critically revisits classic creativity theories that emphasize divergent thinking as the core mechanism, positing instead that creative idea generation and subsequent valuation are inseparable components of a unitary cognitive operation. While divergent thinking—the generation of many ideas—remains important, without a robust evaluative system, it becomes inefficient or chaotic. Subjective valuation thus acts as a gatekeeper, guiding attention and cognitive resources to ideas with the highest perceived potential utility or originality, aligning with a cost-benefit analysis model.

Importantly, this domain-general framework can explain the transferability of creativity across fields. Since subjective valuation processes employ similar neural circuits and cognitive strategies irrespective of content, individuals adept in these evaluative operations can flexibly apply their creative capacities to language, art, science, or technology. This insight challenges educational paradigms that compartmentalize creativity training, advocating for approaches that instead strengthen the underlying evaluative scaffolding.

Moreover, the integration of subjective valuation into creativity models helps reconcile the tension between randomness and control in creative thought. While some models emphasize serendipity and stochastic exploration, Battistello et al.’s results highlight how controlled, value-guided selection is fundamental to progress beyond mere novelty. Creativity thus emerges as a balance between spontaneous generation driven by associative processes and deliberate selection steered by subjective valuations that weigh emotional, aesthetic, and pragmatic considerations.

From a methodological perspective, the study’s combination of neuroimaging and computational techniques provides a sophisticated lens to dissect the temporal dynamics of creativity. The authors demonstrate that valuation signals peak at critical decision points during idea refinement, underscoring the iterative nature of creative thought. This challenges simpler linear models of creativity and advocates for conceptualizing the creative process as a fluctuating negotiation between evaluation and elaboration over time.

In sum, Battistello et al.’s research constitutes a paradigm shift by positioning subjective valuation at the heart of creative cognition. It offers a robust neurocognitive explanation for how individuals navigate the vast and often ambiguous creative landscape, emphasizing a conserved, domain-general evaluative system that orchestrates divergent and convergent thinking phases. These insights pave the way for future interdisciplinary research and practical applications, ranging from cognitive enhancement to artificial intelligence systems that emulate human-like creativity by embedding flexible valuation algorithms.

Intriguingly, the findings may also illuminate clinical conditions where creativity is impaired or atypical, such as in certain psychiatric disorders. Aberrant valuation mechanisms could explain reduced or altered creative expression, offering potential targets for therapeutic intervention. Conversely, understanding how valuation drives creativity might facilitate harnessing creative strengths in neurodivergent populations, fostering inclusive environments that recognize diverse cognitive styles.

Looking forward, the authors suggest key directions for extending their work. These include exploring how subjective valuation interacts with motivational and emotional systems over longer temporal scales, elucidating developmental trajectories of valuation networks in creativity, and leveraging real-world creative tasks outside laboratory settings. Moreover, integrating neural data with rich qualitative analyses of creative products promises to deepen understanding of how valuation informs not just quantity but quality and impact of creative output.

This study’s contributions resonate profoundly as society increasingly seeks to nurture innovation in complex and interdisciplinary domains. By uncovering universal evaluative mechanisms that underlie creative thinking across fields, it provides a conceptual and empirical blueprint for fostering human creativity in a holistic and scientifically grounded manner. As science unfolds the cognitive algorithms at play, the age-old mystery of creativity inches closer to a tangible, mechanistic understanding, promising exciting advances for education, technology, and mental health.

Subject of Research: The neural and cognitive role of subjective valuation as a domain-general process in creative thinking.

Article Title: Subjective valuation as a domain-general process in creative thinking.

Article References:
Battistello, G., Moreno-Rodriguez, S., Volle, E. et al. Subjective valuation as a domain-general process in creative thinking. Commun Psychol 3, 108 (2025). https://doi.org/10.1038/s44271-025-00285-8

Image Credits: AI Generated