Irrational human thinking may be the missing key to creative AI

A new study suggests that human irrationality, long considered a cognitive flaw, may be the key to unlocking machine creativity. Published in Frontiers in Artificial Intelligence, the paper titled “Irrationality in Humans and Creativity in AI” by Olha Sobetska of the Vrije Universiteit Brussel argues that certain irrational decision-making patterns, such as the conjunction fallacy, are not merely errors in reasoning but potentially adaptive mechanisms that support creative insight.

Rather than reinforcing existing AI paradigms centered on logic, optimization, and consistency, the study proposes a hybrid cognitive model that integrates both rational and irrational dynamics. The ultimate goal: to design AI systems that mimic the flexible, context-sensitive, and inventive qualities of human thought.

Can irrationality really enhance AI creativity?

The study reinterprets conjunction fallacy - a well-documented cognitive bias in which individuals rate the probability of a compound event as higher than that of a single component event, violating classical probability theory. This phenomenon, the study argues, should not be dismissed as mere irrationality. Instead, it can reflect context-sensitive reasoning that aligns better with real-world problem-solving than rigid statistical norms.

By analyzing how even highly educated participants persistently fall into this so-called fallacy under certain conditions, the paper highlights a fundamental tension: real-life decision-making often operates outside idealized mathematical frameworks. These deviations, rather than indicating flawed thinking, may reflect cognitive strategies adapted for complex, uncertain environments - hallmarks of creative behavior.

This line of reasoning extends to artificial intelligence. If current generative models lack creativity, it could be due to their overreliance on statistical pattern recognition and insufficient modeling of the nonlinear, sometimes illogical thought processes that fuel human creativity. The study posits that AI could benefit from embracing “rational irrationality” - the productive, context-driven deviations that humans exhibit.

How do methodological fallacies hinder true understanding in science?

To reinforce this argument, the study critiques how scientific methods themselves often impose flawed interpretations on complex phenomena. It identifies a broader pattern of methodological fallacies, including the ecological fallacy in sociology, the misuse of p-values in biomedicine, and generalizations in linguistics that overlook cultural and contextual specificity.

These examples illustrate a core issue: applying methods without regard for context can produce misleading conclusions. In biomedicine, for instance, overreliance on p-values can obscure clinically meaningful effects, while in linguistics, attempts to universalize grammar can erase valuable variation. These oversights are not simply technical - they reflect a deeper failure to align tools with the nuanced nature of the subject matter.

This critique is especially relevant for AI, which often relies on datasets and optimization criteria that flatten context. If AI systems are trained to simulate “rational” behavior based on idealized datasets, they may miss the creative leaps that occur when humans violate norms in purposeful or serendipitous ways.

By re-evaluating such methodological assumptions, the study challenges developers and researchers to build AI that can function effectively in messy, real-world environments. It suggests that the cognitive diversity of human reasoning, including its inconsistencies, should be embraced, not engineered out.

What does a neuro-cognitive model of AI creativity look like?

The study builds toward a neurocognitively informed model of creativity, combining insights from cognitive psychology and neuroscience. It introduces a revised definition of creativity as a dynamic balance between two forces: the rational, focused processing associated with the brain’s Executive Control Network (ECN), and the chaotic, spontaneous flow of ideas linked to the Default Mode Network (DMN). A third player, the Salience Network (SN), mediates the switching between these states.

This brain-based model of creativity aligns with the dual-processing framework: convergent thinking narrows focus to solve structured problems, while divergent thinking enables the exploration of unconventional solutions. The study argues that AI systems should replicate this interplay. Transformer models like GPT, it notes, have taken initial steps in this direction through context-sensitivity and randomness parameters such as “temperature,” which influence the diversity of generated outputs.

Beyond technical tuning, the study points to the promise of skill-mix algorithms that simulate domain-specific semantic fluency, traits closer to human creative cognition. These mechanisms can yield more surprising and context-appropriate outputs, hinting at AI’s potential to move beyond mere mimicry toward authentic creative behavior.

Yet, significant gaps remain. AI lacks intuitive leaps, embodied understanding, and emotional nuance - all essential components of human creativity. The paper concludes that enhancing machine creativity depends not on engineering greater logic but on capturing the full spectrum of human cognition, including its flaws, biases, and irrational brilliance.