How Artists Think — The Evidence

Written by the ENSI Foresight Division on a library of 175 primary documents — studies, meta-analyses, corpus analyses and neuroimaging experiments — downloaded and indexed in the Original Thinking library.

The argument, before the list

The received story about artists is a story about temperament. On this account, some people are born with a strange sensibility — they see more, feel more, tolerate more chaos — and art is what that sensibility does when left alone. The story is convenient for everyone: it lets artists claim a mystery, lets scientists dismiss one, and lets educators file “creativity” under things that cannot be scheduled. The laboratory record, assembled here from 175 primary documents, says the story is wrong in a specific and useful way. Artistic thinking is not a temperament. It is a trainable cognitive stack — a set of distinct, measurable operations in perception, process and cognition that trained artists demonstrably run differently from everyone else, and that untrained people begin to run differently after training.

The evidence is not anecdote. It is drawing-accuracy experiments and eye-tracking traces (Cohen and Bennett’s classic isolation of misperception as the main source of drawing error; Cohen’s finding that gaze frequency predicts drawing accuracy), longitudinal neuroimaging of art students whose brains restructure across a course of training (Schlegel and colleagues’ “The Artist Emerges”), fMRI of jazz musicians and freestyle rappers generating new material inside a scanner (Limb and Braun; Liu and colleagues), corpus analyses of 150 Hollywood films across 70 years (Cutting, DeLong and Nothelfer), network-science mapping of semantic memory in high-original individuals (Kenett, Anaki and Faust), and a meta-analytic literature — on incubation (Sio and Ormerod), on personality (Feist), on the prefrontal geography of idea generation (Gonen-Yaacovi and colleagues) — that has quietly converged while the temperament myth held the stage. Where the artist’s advantage has been looked for in the eye itself, it has not been found: Perdreau and Cavanagh’s test of the old claim that artists “see their retinas” came back negative. The advantage lives higher up — in trained attention, trained selection, trained process. Which is precisely what makes it transferable.

Why should anyone outside the studio care? Because originality has become the binding constraint on the systems that matter. The bibliometric record of science itself — angle twelve of this library — shows papers and patents becoming measurably less disruptive over six decades (Park, Leahey and Funk), novel work systematically under-rewarded at first (Wang, Veugelers and Stephan’s bias-against-novelty analysis), while the work that ultimately lands hardest is the work that injects atypical combinations into conventional cores — Uzzi and colleagues’ finding across 17.9 million papers. Science, in other words, has a documented originality shortage and a documented originality premium, simultaneously. Any institution that could reliably increase its supply of original cognition would be buying the scarcest input in the knowledge economy at exactly the moment machine systems are commoditising the conventional kind.

And there is a population that has been training original cognition, deliberately and for centuries: artists. The transfer evidence makes the connection more than poetic. Root-Bernstein and colleagues’ studies of Nobel laureates found them many times likelier than average scientists to sustain serious arts and crafts avocations — a result extended in their 2019 PNAS-line study of STEMM professionals, where arts, crafts and design practices track scientific achievement. Ramón y Cajal, the founding draughtsman of neuroscience, is the canonical case — a recent essay in the library traces how his drawing practice shaped what he could see down a microscope. The correlational nature of this evidence has limits, and this report will be honest about them — the OECD’s Art for Art’s Sake review found the causal case for far transfer weak — but the direction of the signal is consistent: the people who change science disproportionately think like artists on the side.

So this report does something deliberately unromantic. It treats “how artists think” as an empirical question with an answerable structure, and ranks what the evidence actually supports. At the top sit findings with replicated, objective, convergent support — trained perception, the filmmaker’s command of attention, the oscillating architecture of the creative process. In the middle sit the cognitive mechanisms — far association, improvisation as a releasable brain state, problem finding, combinatorial imagination. At the bottom sit the findings whose effects are large but whose causality is weakest — personality, embodiment, and transfer itself. Nothing here requires belief in genius. Everything here is a difference someone measured.

The reframe to hold through all ten findings is this: what artists possess is not a gift but a stack — trained seeing, trained problem-formulation, trained oscillation between generating and judging, a loosened associative network, a practised ability to switch off self-censorship on demand, and a body and studio arranged as thinking instruments. Every layer of that stack has been measured. Several have been trained in non-artists inside controlled studies. That is the practical meaning of the evidence: originality is a capability, not a lottery — and the artist is its best-documented working model.

The findings in brief

• Artists see differently — and the seeing is trained, not given. Drawing accuracy is limited by perception, not the hand; artists’ eye movements, salience resistance and shape perception all differ measurably; training changes the brain (Cohen & Bennett · Chamberlain & Wagemans · Perdreau & Cavanagh · Schlegel).

• Filmmakers are running experiments on attention — and winning. Edit blindness, gaze synchrony, event segmentation and 70 years of corpus evolution show cinema is applied cognitive science (Smith · Hasson · Zacks · Cutting · Cohn).

• The creative process is an oscillation between generative and evaluative modes, visible in the brain and in artists’ own accounts (Ellamil · Sowden, Pringle & Gabora · Beaty · Botella).

• Original ideas come from far associations. High-original people carry measurably more flexible semantic networks, and originality rises with semantic distance and with time-on-task (Kenett · Benedek & Neubauer · Beaty & Silvia · Green).

• Improvisation shows originality is a releasable brain state — self-monitoring down, self-expression up — and improv training raises originality in ordinary teenagers (Limb & Braun · Liu · Hainselin).

• Artists find problems before they solve them. Problem construction measurably improves creative outcomes; studio process is discovery-shaped, not execution-shaped (Reiter-Palmon · Botella · Tversky · Scotney).

• Analogy, metaphor and blending are the combinatorial engines — the same machinery Dunbar filmed in world-class science labs (Gentner · Holyoak · Fauconnier & Turner · Dunbar · Uzzi).

• The original personality exists and is measurable — openness to experience is its strongest marker, split revealingly between artistic and scientific forms (Feist · Kaufman · Barron · Kyaga).

• The body is part of the thinking. Incubation works, walking boosts ideation, choreographers think with their limbs, and groups generate what no member holds (Sio & Ormerod · Oppezzo & Schwartz · Kirsh · Sawyer).

• Artistic originality transfers — with honest limits. Nobel-laureate avocations and STEMM correlations are strong signals; causal far-transfer evidence remains weak (Root-Bernstein · OECD · Schellenberg).

How this report is organised

The ten findings are ranked by strength and convergence of evidence, judged on four tests: whether the core result has been replicated; whether independent methods converge on it (behavioural experiment, eye-tracking, neuroimaging, corpus analysis, meta-analysis); whether the measures are objective rather than self-reported; and whether anything causal — a training study, an intervention — anchors it. A finding built on decades of replicated experiments with objective measures outranks one built on a single elegant fMRI study; a mechanism confirmed by meta-analysis outranks a beautiful theory; correlational biography, however striking, ranks last no matter how large its effect sizes. Each finding follows the same discipline: the claim in one line, the actual studies from the library, an explicit account of what the evidence does not show, and what a scientist, founder or policymaker should do with it. The library’s fifteen angle indexes are the bibliography; every study named below is in them.

1. Artists see differently — perception is trained, not given

The best-documented difference between artists and everyone else is perceptual: artists have retrained what they see, and the retraining is measurable at every level from eye movement to brain structure.

The foundation stone is Cohen and Bennett’s 1997 study, bluntly titled “Why can’t most people draw what they see?” By experimentally isolating the candidate failure points — misperception of the object, inability to make good representational decisions, misperception of one’s own drawing, motor incoordination — they showed that the main obstacle is the first one: people draw badly because they see conventionally. The hand is largely innocent; the eye is guilty. Chamberlain and Wagemans’ comprehensive review of the genesis of drawing errors traces the same conclusion across the perception, memory and motor stages of drawing, and the Kozbelt line of work — updated and replicated in Chamberlain, Drake, Kozbelt and colleagues’ “artists as experts in visual cognition” study — shows that trained artists outperform non-artists on visual tasks well beyond drawing itself. Robles, Bies, Lazarides and Sereno’s 2022 Scientific Reports study closes the loop in the modern era: veridical shape perception measurably tracks drawing ability.

The eye-tracking evidence shows how the trained eye works. Cohen’s 2005 study found that trained artists shift gaze between object and drawing far more frequently than novices — and that gaze frequency itself predicts drawing accuracy, as though expertise consisted partly of refusing to let memory adulterate the percept. Park, Williams and Chamberlain found artists’ global-to-local saccade ratios while drawing differ systematically from non-artists’. Koide and colleagues’ PLOS ONE study delivers the sharpest formulation: when experts view abstract paintings, their fixations are driven far less by bottom-up visual salience than novices’ — the trained eye sees top-down, going where its owner’s questions direct it rather than where the stimulus shouts. Francuz and colleagues found the same expertise signature in fixation patterns during compositional judgement.

Two further results give the finding its edge. First, Perdreau and Cavanagh’s direct test of the romantic hypothesis — that artists somehow access a raw, less-interpreted “retinal” image — came back negative: artists’ advantage is not low-level vision. It is learned control over attention, selection and constancy, which is exactly why it is trainable rather than congenital. Second, Schlegel and colleagues’ longitudinal NeuroImage study followed art students across their training and watched neural structure and function change as drawing skill grew — the artist, as their title has it, emerges. Perception here is not a window but an instrument, and instruments can be re-machined.

The neuroaesthetics canon shows what this trained seeing plugs into on the receiving side. Vessel, Starr and Rubin found that the most intensely moving aesthetic experiences engage the default-mode network — art reaching the systems of the self, not merely the eyes; Ishizu and Zeki located the experience of beauty, across visual and musical modalities, in medial orbitofrontal activity; and Chatterjee and Vartanian’s aesthetic triad — sensory-motor, emotion-valuation, meaning-knowledge — maps the full circuit an artwork traverses. Massaro and colleagues show viewer expertise reshaping even how paintings are visually explored. Perception in art is never just optics; it is optics wired to valuation and selfhood — which is why retraining it changes more than drawing.

What the evidence does not show: it does not show that artists have better eyesight, faster visual processing, or any general perceptual superiority — the advantage is specific to trained operations of attention and selection. Nor does the correlational strand (shape perception tracking drawing skill) by itself prove direction; the longitudinal and training evidence carries the causal weight.

The “so what” is the most direct in this report, because the tool is sitting on the desk. Fan, Bainbridge, Chamberlain and Wammes’ Nature Reviews Psychology review positions drawing as a versatile cognitive tool that changes what its user perceives, remembers and communicates — a claim with Cajal as its historical proof. A scientist or founder who learns observational drawing is not acquiring a hobby; they are running the best-validated perceptual retraining programme on record. If your work depends on noticing what the conventional eye smooths over — anomalies in data, unmet needs in a market, the detail everyone’s schema deletes — the artists’ evidence says: noticing is a skill with a syllabus.

2. Filmmakers run experiments on attention — film is applied cognitive science

A century of filmmaking craft encodes a working science of human attention — and when cognitive scientists finally tested it, the filmmakers were right.

This finding ranks second because its evidence is unusually convergent: eye-tracking, fMRI, behavioural change-detection and corpus analysis all land on the same conclusion from independent labs. The founding document is Hasson and colleagues’ 2008 neurocinematics paper, which put viewers of different films in a scanner and measured intersubject correlation — how similarly different people’s brains respond, moment by moment, to the same footage. The result: films differ dramatically and measurably in how tightly they control viewers’ brains. Directorial control is not a metaphor. It is a quantity, and some directors have more of it.

Tim Smith’s programme at Birkbeck explains the mechanism. His attentional theory of cinematic continuity formalises what editors discovered by feel: continuity editing works because it exploits the machinery of visual attention, placing cuts where the visual system is already committed to a saccade, a motion, an expectation. His edit-blindness eye-tracking study with Henderson showed the astonishing consequence — viewers simply fail to see cuts that follow the continuity rules, missing splices in the visual stream that a naive theory of perception says should be jarring. His later work with Martín-Portugués showed how match-action motion and audio timing manufacture global change blindness across cuts. Editors, in other words, spent a century converging on the parameters of change blindness decades before psychology named the phenomenon — a craft tradition that discovered real cognitive law by iteration, the way medieval builders discovered statics.

The corpus evidence shows the discovery process operating at industrial scale. Cutting, DeLong and Nothelfer analysed 150 Hollywood films across 70 years and found shot lengths evolving, decade by decade, toward the 1/f temporal rhythms characteristic of natural human attention — an entire industry performing a slow, unplanned gradient descent onto the attention dynamics of its audience. Cutting’s companion corpus work maps how filmmakers construct narrative space shot by shot. Zacks and colleagues’ fMRI work shows viewers’ brains segmenting narrative film into events precisely at the boundaries editors chose, and Magliano and Zacks show continuity editing shaping those segmentation networks during commercial film viewing. Cohn’s visual narrative grammar completes the picture from the linguistics side: sequential visual storytelling has a hierarchical constituent structure — a real grammar, extended in his later work from static sequences to film — which is why some cut orders parse and others do not; his tutorial paper turns the grammar into a practical toolkit for analysing any storyboard or sequence. Smith’s psychocinematics chapter closes the methodological loop, using eye-tracking to test film theory’s own claims about how directors steer the gaze — a hundred years of critical assertion suddenly exposed to falsification, and much of it holding.

What the evidence does not show: that filmmakers hold this knowledge explicitly — the craft knows things its practitioners cannot state, which is exactly what makes it interesting as a model of embodied expertise. And Loschky and colleagues’ “tyranny of film” study draws an important boundary: strong gaze synchrony does not guarantee shared understanding — where viewers look converges far more than what they comprehend. Attentional control is not semantic control.

The “so what” runs in two directions. For anyone who designs experiences that must pass through human attention — products, interfaces, briefings, lectures — cinema is a validated engineering literature, not an entertainment. Its principles (guide the eye before the cut; ride prediction rather than fighting it; segment events where the mind already segments) are testable and tested. For research strategy, the deeper lesson: mature craft traditions are unread datasets. Editing anticipated change-blindness research; drawing instruction anticipated perceptual-expertise research. A foresight-minded institution should ask which of today’s craft communities — game designers, prompt engineers, standup comedians — are currently sitting on the next such body of pre-scientific cognitive law.

3. The creative process is an oscillation between generating and evaluating

Creation is not one mental mode but a disciplined alternation between two — and the brain dissociates them cleanly enough to see in a scanner.

The anchor study is Ellamil, Dobson, Beeman and Christoff’s 2012 NeuroImage experiment, which put participants through cycles of designing book covers — generate, then evaluate, then generate again — inside an fMRI scanner. Generation preferentially engaged medial temporal regions associated with associative memory and novel combination; evaluation recruited executive regions together with default-mode regions, suggesting creative judgement is not cold quality control but a hybrid of analysis and imaginative simulation. The two modes are neurally dissociable, and the creative process consists of moving between them — not of occupying some single “creative state”.

The behavioural and theoretical record converges. Sowden, Pringle and Gabora’s dual-process account describes creation as shifting between associative and analytic modes, with skill lying in the timing of the shifts. Botella, Zenasni and Lubart’s study of art students’ own process accounts finds the same shape from the inside: real artistic process is iterative and looping — idea, test, judgement, re-entry — rather than the tidy linear pipeline of textbook stage models. Beaty and colleagues supply the network-level mechanism: their Scientific Reports fMRI study found that creative idea production is supported by coupling between the default network and the executive control network — two systems textbook neuroscience once treated as antagonists, working in alternating cooperation — a result consolidated in Beaty’s 2019 review of the creative brain’s network architecture. De Pisapia and colleagues found the same default-executive cooperation in professional artists planning a real artwork, which matters: the oscillation is not an artefact of laboratory tasks. It is visible even inside a single act of making — Miall, Nam and Tchalenko’s fMRI study of portrait drawing, from the team behind the celebrated eye-and-hand studies of the painter Humphrey Ocean, catches the drawing brain cycling between looking, deciding and executing rather than running one continuous programme.

Chrysikou and colleagues’ matched-filter hypothesis adds the report’s most counterintuitive mechanism: prefrontal cognitive control is not simply good — it is task-matched. High control serves well-defined problems; low control measurably benefits open-ended generation, which is why deliberate effort so often strangles the very ideas it is trying to produce. And Gonen-Yaacovi and colleagues’ ALE meta-analysis of 34 fMRI studies locates creativity’s generative and combinatorial operations across rostral and caudal prefrontal regions — solid meta-analytic ground under the claim that generation and evaluation are distinct computations, not moods.

What the evidence does not show: any single “creativity centre”, and emphatically not the folk right-hemisphere story. Boccia and colleagues’ meta-analysis finds musical, verbal and visuospatial creativity resting on partly domain-specific networks — creation is an orchestration, differently cast per domain.

The “so what” is organisational as much as personal. Most knowledge work runs generation and evaluation simultaneously — the meeting that brainstorms and critiques in the same breath, the writer who edits each sentence as it appears — which the matched-filter and dual-process evidence identifies as running both modes at once and doing each badly. Artists’ studio practice institutionalises the separation: sketch phases where nothing is judged, crit sessions where everything is. A lab or company can copy this directly — separate generative sessions from evaluative ones in time, place and even personnel, and treat the switch as the skill to train. The oscillation is the process; scheduling it is management.

4. Original ideas come from far associations — and the network is measurable

Highly original people carry differently structured semantic memories — more flexible, better connected across distant regions — and originality rises measurably with associative distance.

This is the mechanism-of-record for where new ideas come from, and its modern evidence is unusually elegant. Kenett, Anaki and Faust applied network science to the semantic memory of low- and high-creative individuals and found the high-creative networks measurably more flexible and interconnected — shorter paths between far-flung concepts, less rigid clustering. An original mind is not a bigger warehouse; it is a better-connected graph, in which “remote” ideas are simply less remote. This gives structural teeth to Mednick’s old associative theory — that creative individuals have flat rather than steep associative hierarchies, so unusual associates are nearly as available as obvious ones — which Benedek and Neubauer tested directly and refined: creative people’s advantage lies substantially in more effective search through associative memory, not merely a different gradient.

The process evidence shows far association operating in time. Beaty and Silvia’s serial-order work documents that ideas become measurably more creative the longer one generates — the first answers are everyone’s answers; originality lives past the point where most people stop. Beaty, Silvia, Nusbaum, Jauk and Benedek’s Memory & Cognition study shows divergent production riding on both associative and executive processes — the loose network plus the disciplined search of it — and Benedek’s work in Intelligence maps how executive functions serve both intelligence and creativity, against any romantic opposition between the two. On the neural side, Green and colleagues found that semantic distance in analogical reasoning parametrically engages frontopolar cortex — the brain treats far connection as its own operation. Kounios and Beeman’s insight programme — the Aha! moment and their Annual Review synthesis — shows solutions-by-insight have distinct neural signatures, with brain states preceding a problem predicting whether it will be solved by insight or analysis. Preparation shapes originality before the problem arrives. And Beaty and Johnson’s SemDis platform now scores associative originality automatically as semantic distance — the construct is measurable enough to be computed.

What the evidence does not show: that divergent-thinking scores are destiny. The library’s 2022 review of the Torrance Tests lays out the predictive-validity debates honestly — test-measured divergence correlates with, but does not guarantee, real creative achievement, which needs domain skill, motivation and opportunity (Amabile’s componential model, from the foundations angle, is the standing corrective).

The “so what”: originality has a substrate you can build. Feed the network heterogeneous material — distant fields, unshared experiences, deep non-work domains — because association can only connect what is stored. Then work the search: persist past the first wave of ideas (the serial-order effect is a free lunch almost nobody eats), and measure output by semantic distance, which SemDis-class tools now make practical for teams. A founder’s differentiated insight and a scientist’s atypical combination are, mechanically, the same event: a traversal between regions of the graph that competitors’ graphs do not connect. And because Kounios and Beeman show pre-problem brain states predicting insight, the preparation is not metaphorical — the mood, attention and expectation a person carries into a problem are part of the solving machinery, and can be set deliberately.

5. Improvisation shows originality is a releasable brain state

When trained improvisers generate in real time, the brain measurably reconfigures — self-monitoring down, self-expression up — and the state can be trained into ordinary people.

The seminal document is Limb and Braun’s 2008 PLOS ONE study, which put professional jazz musicians on a custom keyboard inside an fMRI scanner and compared improvisation against memorised performance. Improvisation came with a characteristic prefrontal reconfiguration — broadly, a retreat of the dorsolateral regions associated with deliberate self-monitoring and control, alongside engagement of medial prefrontal regions associated with self-expression. Liu and colleagues then found the same dissociated medial-versus-dorsolateral prefrontal pattern in an entirely different art form, freestyle rap, comparing improvised with rehearsed lyrics. Two genres, two labs, one signature: the generative state is partly a release — a temporary standing-down of the very supervision that ordinary cognition works so hard to maintain. This is the matched-filter hypothesis (finding 3) caught live: less top-down control precisely when the task is open-ended generation.

The surrounding evidence widens the base. Saggar and colleagues’ Pictionary-style fMRI study of improvised drawing found cerebellar-cortical dynamics supporting spontaneous figural creativity — recruiting the brain’s motor-automation machinery for idea generation, a hint that fluent generation behaves like a trained skill rather than deliberate reasoning. Norgaard and colleagues’ study of artist-level jazz improvisers maps the functional network connectivity of the practised improvising brain, and Arkin, Przysinda and Loui find structural grey-matter correlates of improvisational creativity — the state leaves a trace, as states that are trained do. Pressing’s foundational cognitive model of improvisation explains what the years of practice build: automated generative structures that free attention from execution — the release is affordable only because the underlying craft has been drilled to the point of costing nothing. And the crucial causal anchor comes from outside music: Hainselin and colleagues showed that an 11-week improvisational-theatre programme raised divergent-thinking originality and flexibility in ordinary teenagers — the state is not a property of elite musicians but a trainable disposition, teachable by curriculum. The applied-improvisation literature in the library explains what such training actually rehearses: the theatre’s “Yes, and” rule is deferral of judgement made into a bodily reflex — acceptance first, evaluation later — which is the generative half of finding 3’s oscillation, installed as habit.

What the evidence does not show: the neuroimaging samples are small and expert-heavy, as first-generation scanner studies are; the precise prefrontal geography varies across studies and tasks; and “deactivation” is a simplification of a reconfiguration that differs between musical and verbal improvisation. The honest claim is not a single switch in the head, but a replicated family of state changes with the same functional meaning: supervision loosens, generation flows.

The “so what”: most professional environments are engineered to keep the monitoring system permanently on — evaluation, status, audit. The improvisation evidence says original output needs sanctioned intervals where it is off, and that the off-switch strengthens with practice. Improv classes for scientists and founders are not team-building whimsy; they are state training with a controlled study behind them. The practical designs follow directly: regular low-stakes generative sessions whose output is explicitly unjudged; warm-up rituals borrowed from performers; and a personal practice — musical, verbal, physical — that rehearses the release itself.

6. Artists find problems before they solve them

The most distinctive move in the artistic process happens before solving begins: original creators treat the problem itself as the thing to be discovered, and problem construction measurably improves creative outcomes.

This is the oldest process finding in the library’s lineage — the problem-finding tradition that began with Getzels and Csikszentmihalyi’s studies of art students, carried forward in the library by its modern heirs. Reiter-Palmon and Murugavel’s review of problem construction consolidates the experimental record: when people actively construct and reformulate a problem before solving — rather than accepting it as given — their solutions are reliably more original and more effective, and problem-construction ability tracks creative performance across studies. Botella, Zenasni and Lubart’s study of art students’ own creative process confirms the shape from the practitioner’s side: the early phases of artistic work are dominated by searching, reframing and defining, not executing. The artist’s question is not “how do I solve this?” but “what is actually worth making?” — and the evidence says the quality of the eventual answer is substantially decided there.

The surrounding studies show how the finding operates in the wild. Tversky’s synthesis of the sketch-cognition programme documents designers and architects discovering problems in their own sketches — drawing something, then seeing in the drawing relations and possibilities they did not knowingly put there; the sketch is a problem-finding instrument, not a record of a finished thought. Glăveanu and Lahlou’s subjective-camera study — head-mounted cameras on working craftspeople — catches the same continuous renegotiation between maker and material in real studio practice. Scotney and colleagues add a structural result: creative inspiration disproportionately crosses domains in the early, idea-finding phase of the process — the problem-finding window is precisely when far material (finding 4) enters. And Dunbar’s in-vivo studies of molecular biology laboratories show the scientific analogue: the labs that generate discoveries treat anomalies not as noise to be explained away but as problems to be adopted — problem finding as institutional reflex.