The Mind Already Knows What Complexity Is

New research in cognitive science suggests that the brain represents complexity as a single unified quantity across all sensory domains - and what that means for how we teach, practise, and understand creative thinking.

A paper published in Nature Human Behaviour earlier this month by Tal Boger and Chaz Firestone at Johns Hopkins University makes a claim that, on the surface, sounds almost too tidy to be true: complexity is a unified cognitive kind. Across eleven experiments involving 1,500 participants, the researchers found that the human mind does not treat the complexity of a visual shape, the complexity of a melody, the complexity of a letter string, and the complexity of a mathematical expression as four separate problems. It treats them as one. Reward associations transfer across domains. The intrusion is automatic. Individual aesthetic preferences - whether someone finds simple shapes pleasing or elaborate ones - predict how they respond to unrelated domains, such as melody or texture. The full study is available in open access.

This is not, on reflection, so surprising - but it is rarely stated so precisely, and the precision matters. What Boger and Firestone are describing is a domain-general representational language, something the brain appears to use to encode information density regardless of the medium in which that density appears. The implication is that when you train a person's sensitivity to complexity in one register, you are not merely teaching a domain-specific skill. You are exercising a shared cognitive capacity that reaches into every other domain simultaneously.

The history of creativity research has generally moved in the opposite direction from this finding. Since the mid-twentieth century, the dominant tendency has been to fractionate. Guilford separated divergent from convergent thinking. Gardner proposed multiple intelligences, each relatively autonomous. Domain specificity became almost an axiom: expertise in one field was held to transfer very little to another, and creativity in music was cordoned off from creativity in mathematics or verbal reasoning. There were good reasons for this fractionation - it pushed back against vague, unmeasurable claims about general creative genius - but it also produced a generation of training programmes that assumed you could build "creativity skills" in isolation from each other and from the sensory fabric of thought.

What Boger and Firestone's data suggest is that there may be a layer beneath domain expertise at which complexity - understood as information density, as the opposite of predictability, as the degree to which a pattern resists compression - is processed by a single representational system. The researchers describe it as consistent with "a shared representational language across cognitive domains." This does not invalidate domain expertise. It does, however, suggest that exposure to genuine complexity in any domain may recalibrate a threshold that operates everywhere. A person who has spent months reasoning about intricate melodic structure may arrive at a mathematical problem or a conversation about abstract ethics with something already tuned.

Arthur Koestler's concept of bisociation - the momentary connection of two normally separate planes of thought that produces a creative act - has always implied something structurally similar, even if Koestler never had neuroimaging or reaction-time data behind him. Bisociation works precisely because the mind is capable of recognizing structural kinship between domains that have no surface resemblance. The moment of connection is not between two sets of facts; it is between two complexities that turn out to rhyme. Spreading activation theory, as developed by Collins and Loftus in the 1970s, described the mechanism by which activating one semantic node radiates outward through a conceptual network - but it was still largely modelled within a single domain. Boger and Firestone are pointing at something that may cross domain boundaries entirely, not by analogical inference but by direct representational transfer.

The practical consequence of this finding for language education is worth sitting with. Advanced language learners plateau, and the diagnosis of why they plateau tends to focus on vocabulary gaps or grammar fossilization. These are real, but there is another possibility: that the learner's appetite for and sensitivity to complexity has itself dulled. Predictable topics, predictable prompts, predictable genre conventions - the standard apparatus of proficiency testing - do not exercise the complexity-detection mechanism. A person trained exclusively on structured academic paragraphs and IELTS cue cards is not being exposed to the kinds of dense, unresolvable, compositionally intricate material that would keep that mechanism active. The consequence is not merely that their writing sounds flat. It is that their capacity for conceptual surprise, for noticing when an idea is genuinely non-trivial, may be quietly contracting.

There is a further issue that the Boger-Firestone finding brings into focus. Generative AI systems produce output that is, statistically, smooth. The outputs of large language models are optimised, at the level of training objective, to be probable - which is to say, to be low-complexity in exactly the sense that information theory would predict. A sentence generated by a language model tends to be the sentence that was most likely given its context. Aesthetic and conceptual density, which require selecting the less probable, the more surprising, the formally complex, are not what these systems maximise. When learners use AI to draft their writing, read AI-generated summaries of complex texts, or receive AI-mediated feedback, they are systematically exposing themselves to a diet that is lower in information density than the texts those systems were trained on. The Boger-Firestone findings suggest this has implications beyond writing quality. If reward associations to complexity transfer automatically across domains, then sustained exposure to low-complexity input may recalibrate downward the threshold at which a person finds simplicity rewarding - and simplicity is not what intellectual growth, creative production, or advanced language use actually requires.

"Outcomes associated with complexity generalize across diverse stimulus classes." -- Boger & Firestone, Nature Human Behaviour, 2026

The experimental design Boger and Firestone used is worth describing briefly because it clarifies what "automatic" means in this context. In one set of conditions, participants performed a task in which rewards were associated with either simple or complex stimuli in one domain; when they then encountered stimuli from a completely different domain, their responses showed that the reward association had already transferred, without any instruction to generalise. This is not analogical reasoning. The person did not consciously think: this melody resembles that shape, therefore I should respond to it similarly. The transfer happened below the level of deliberate inference. The researchers describe the intrusion as occurring on task-irrelevant judgements - meaning that even when people were not asked to evaluate complexity, their prior associations were influencing their responses. This is what domain-general representation looks like in practice.

For those of us working in creativity training, the word "automatic" here is the operative one. Training that only happens when a person consciously applies a technique is useful but limited. What you actually want is recalibration - a change in the baseline at which the cognitive system operates, not just an additional skill that can be called upon when remembered. Random Abstractions, for instance, works at the level of noticing structural kinship between domains that have no surface connection. The activity at grandomastery.com/abstractions asks learners to compare two unrelated abstract nouns not by their dictionary definitions but by the underlying logic of what each one does, how it moves, what it resists. The cognitive work required is exactly the kind of cross-domain complexity matching that Boger and Firestone's experiments reveal to be unified in the mind. What this new research adds is a reason to believe that practising it with abstract nouns has consequences for how the same person hears a melody, reads a mathematical expression, or responds to an unexpected texture. The transfer is not a metaphor. It may be structural.

There is a historical irony in the timing of this finding. The twentieth century - particularly the decades after the Second World War - produced an enormous investment in arts education as a vehicle for general cognitive development. Music lessons, drawing, poetry analysis, close reading of complex prose: these were defended not only on cultural grounds but on the grounds that they trained something transferable. That defence proved difficult to sustain under the pressure of narrowly empirical accountability frameworks, because the transfer was real but the mechanism was invisible. Boger and Firestone are, in a sense, describing the mechanism. The transfer was always domain-general because complexity was always domain-general. The reason a violinist and a mathematician might share a particular quality of attention is not that music and mathematics are secretly the same thing. It is that the mind uses the same representational vocabulary to measure how dense, how surprising, how far from trivial each of them is.

This also has something to say about individual differences in creative output that tend to be attributed, somewhat loosely, to personality. The finding that people who find simple shapes aesthetically pleasing also find simple melodies pleasing is not trivially obvious. Aesthetic preferences are often treated as highly personal and domain-specific - one person loves sparse minimalist design, another loves maximalist visual density, and these seem like independent stylistic choices. What the data suggest is that there is a single underlying dimension of complexity preference that runs through aesthetic responses in general. This matters for creativity training because it means the dimension can be trained. A person's current position on it is not a fixed personality trait. It is a calibration point that responds to experience.

None of this implies that domain expertise is irrelevant. The complexity that a professional musician perceives in a piece of counterpoint is not identical to the complexity a novice perceives, even if both are drawing on the same domain-general representational faculty. Deep knowledge shapes what counts as complex within a domain. What Boger and Firestone are describing operates at a level of abstraction that is prior to domain expertise - a shared metric, not a shared body of knowledge. The relationship between them is something like the relationship between sensitivity to light and the ability to read a particular language: the first is necessary for the second but does not substitute for it.

For educators and trainers who work with advanced language learners, the research opens a question that is worth taking seriously: to what extent does the material we use to build linguistic complexity also build the domain-general complexity-sensitivity that would make learners more responsive to nuance, more tolerant of ambiguity, more capable of generating genuinely non-predictable language? A learner who is regularly exposed to genuinely complex stimuli - not complicated in the bureaucratic sense, but dense with information, resistant to easy compression, containing more than one simultaneous logic - is not merely practising language. They may be tuning a shared cognitive system that influences how they process and produce meaning across every domain in which they operate. That seems worth knowing.

The study is available at nature.com/articles/s41562-026-02502-8. More on bisociation, cross-domain ideation, and structured spontaneity at grandomastery.com.

Source: Boger, T. & Firestone, C. (2026). Complexity is a unified cognitive kind. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02502-8