Intelligence

Abstract

Intelligence is the general cognitive capacity that lets a person reason, solve novel problems, plan, think abstractly, comprehend complex ideas, and learn from experience. In psychometrics it is the construct that explains the positive manifold — the near-universal finding that people who do well on one cognitive test tend to do well on all of them — and that finding, first quantified by Spearman in 1904, motivates the general factor g that sits at the top of every major hierarchical model. This article covers what the construct is and is not, its factorial structure from g through broad abilities to narrow skills, how it is measured and scaled, the neuroscience of individual differences, and the behavioural and molecular genetics of intelligence, closing with the genome-wide and neuroimaging work now defining the field's active front.

Keywords: intelligence, general factor, fluid and crystallized abilities, psychometrics, heritability

Few constructs in psychology are as robustly measured, as strongly predictive, and as persistently misunderstood as intelligence. A working definition endorsed by a broad group of researchers describes it as “a very general mental capability that, among other things, involves the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly and learn from experience” (Gottfredson, 1997). It is not the sum of what a person happens to know, nor a single skill; it is the capacity that makes acquiring skills and knowledge faster and more reliable. The scientific study of intelligence rests on a century of measurement, and its central empirical facts — the positive manifold, the predictive validity of test scores, the substantial and age-increasing heritability of individual differences — are among the most replicated in the behavioural sciences (Deary, 2012).

What Intelligence Is

Intelligence is defined operationally by what intelligence tests measure and theoretically by the structure of individual differences those tests reveal. The two are tied together by a single empirical regularity. When a large sample takes a battery of varied cognitive tasks — vocabulary, matrix reasoning, arithmetic, spatial rotation, memory span — every pair of tasks correlates positively. No one has ever assembled a broad battery in which the correlations vanish or turn negative; this is the positive manifold (Spearman, 1904). Something shared drives performance across otherwise dissimilar tasks, and factor analysis of the correlation matrix extracts that shared variance as a general factor, g.

g is not a thing in the head; it is a statistical summary of covariance among abilities. But it is a remarkably durable one. It emerges whatever the specific tests, provided the battery is broad, and the g estimated from one battery correlates almost perfectly with the g from an independent battery. Its predictive reach is what makes it more than a curiosity: g forecasts educational attainment, job performance (increasingly so as job complexity rises), income, physical health, and life expectancy (Deary, 2012; Gottfredson, 1997). Intelligence is therefore best understood not as a score but as the construct that explains why scores across the cognitive domain cohere and why they matter for life outcomes.

The construct must be distinguished from adjacent ones. It is not knowledge, though the two are correlated; it is not working memory, though working-memory capacity is one of g's strongest correlates; and it is not executive function, though tests of reasoning load heavily on executive control. Intelligence is the higher-order capacity these systems jointly serve.

Types of Intelligence

MeSH classifies Intelligence under Personality (tree F01.752) and lists the narrower descriptors below as its direct subtypes. The single indexed subtype is shown in Table 1. Because MeSH is an indexing classification — a controlled vocabulary for cataloguing the literature — this list reflects how research is filed, not a theoretical claim about how the mind carves into kinds; the psychometric structure in the next section (a hierarchy of g, broad abilities, and narrow skills) is the substantive taxonomy. The subtype below is also not mutually exclusive with general intelligence: emotional intelligence correlates with g while adding domain-specific variance of its own.

The Structure of Intelligence

If g were the whole story, a single number would exhaust cognitive ability. It does not. Above the level of individual tests but below g sit broad abilities — clusters of related skills that correlate more tightly with one another than with the rest of the battery. The most influential distinction, drawn by Cattell, splits g into two broad factors: fluid intelligence (Gf), the capacity to reason and solve novel problems independently of learned knowledge, and crystallized intelligence (Gc), the accumulated knowledge and verbal skill a person has acquired through experience and education (Cattell, 1963). Fluid ability peaks in early adulthood and declines gradually thereafter; crystallized ability is maintained or grows into late life. The dissociation is visible in everyday cognitive aging: an older adult solves an unfamiliar logic puzzle more slowly than a young adult but commands a larger vocabulary.

The first serious challenge to a single g came from Thurstone, who factor-analysed a broad battery and argued that ability resolved not into one general factor but into several relatively independent primary mental abilities — verbal comprehension, word fluency, number, space, associative memory, perceptual speed, and reasoning (Thurstone, 1938). The apparent conflict with Spearman dissolved once those primary abilities were themselves found to intercorrelate, implying a higher-order factor sitting above them; the modern hierarchy is precisely the reconciliation of Spearman's g with Thurstone's group factors. Cattell's two-factor scheme was extended by Horn and then synthesised with Carroll's exhaustive re-analysis of hundreds of datasets into the Cattell–Horn–Carroll (CHC) model, now the consensus framework for test construction. CHC is a three-stratum hierarchy: g at the top (stratum III), roughly a dozen broad abilities beneath it (stratum II), and dozens of narrow, specific skills at the base (stratum I). Table 2 lists several of the broad abilities and what each contributes.

Alternative frameworks reject the primacy of g. Sternberg's triarchic theory distinguishes analytical, creative, and practical intelligence, arguing that conventional tests capture only the first. Gardner's theory of multiple intelligences proposes several largely independent faculties — linguistic, logical-mathematical, spatial, musical, bodily-kinesthetic, and others. These theories broaden the everyday meaning of intelligence, but the proposed factors, where they can be measured, tend to correlate positively — reinstating a general factor the theories set out to dissolve (Neisser et al., 1996). The positive manifold is the fact every structural theory must accommodate.

Figure 1

The Three-Stratum Hierarchy of Cognitive Ability

Measuring Intelligence

Intelligence is measured with standardized batteries — the Wechsler scales and the Stanford–Binet chief among them — that sample several broad abilities and combine the subtests into a full-scale intelligence quotient. The modern IQ is a deviation score: raw performance is ranked against an age-matched norm sample and rescaled to a normal distribution with a mean of 100 and a standard deviation of 15. A score of 130 therefore falls two standard deviations above the mean, at roughly the 98th percentile, and carries the same meaning at any age. This is the province of psychometrics, which supplies the reliability and validity evidence a test must meet.

Good tests are strikingly reliable: full-scale scores show test–retest correlations above 0.9, and rank order is stable across decades. In one landmark demonstration, people retested at age 77 on a test they had first taken at age 11 showed a correlation above 0.6 across 66 years (Deary, 2012). The interactive demonstration below shows how a single deviation score maps onto the normal curve and its percentile.

The first demonstration lets a reader place a score on the IQ distribution and read off its percentile and rarity.

Test scores are not fixed marks of worth; they are estimates, with a standard error, of a person's standing on a broad and consequential dimension of individual difference. Their value lies in that predictive standing, not in any single number.

The Neuroscience of Intelligence

If g reflects a real property of the nervous system, that property should have neural correlates. Early work linked higher intelligence to whole-brain volume, a correlation of about 0.2–0.3 that is real but modest — size is not the mechanism. A more specific account came from functional imaging. Duncan and colleagues found that demanding, g-loaded tasks of different content — verbal, spatial, numeric — converged on a common set of lateral prefrontal and parietal regions, suggesting that g indexes the efficiency of a distributed control network rather than any single structure (Duncan et al., 2000).

This became the Parieto-Frontal Integration Theory (P-FIT), which synthesised dozens of structural and functional studies into a model in which intelligence depends on the flow of information between posterior association cortex, where sensory information is integrated, and prefrontal cortex, where it is evaluated and acted upon, along the white-matter tracts that connect them (Jung & Haier, 2007). On this view individual differences in intelligence arise less from any one brain region than from the integrity and efficiency of the parieto-frontal network as a whole — a proposal consistent with the strong relationship between white-matter integrity, processing speed, and g.

The second demonstration renders the P-FIT network and lets a reader see which nodes a given cognitive operation recruits.

Neuroscience has thus reframed the old question. The issue is not where intelligence lives but how efficiently a widely distributed system communicates — a question that connects psychometric g to the biology of neural conduction and connectivity (Deary, Cox, & Hill, 2022).

Genetics and Heritability

Individual differences in intelligence are substantially heritable. Twin, adoption, and family studies converge on a heritability of roughly 50% averaged across the life span, and — counterintuitively — that figure rises with age, from about 20% in early childhood to 60% or more in adulthood, a pattern known as the Wilson effect (Plomin & Deary, 2015). Heritability increasing with development runs against the intuition that experience should accumulate and swamp genetic influence; the leading explanation is gene–environment correlation, whereby people increasingly select, evoke, and shape environments that match and amplify their genetic propensities.

Heritability is a population statistic, not a personal one, and it is routinely misread. It describes the proportion of variance between people in a particular population and environment that tracks genetic differences; it says nothing about the origin of any individual's ability and does not imply that scores are fixed. The Flynn effect — the substantial rise in raw test scores across the twentieth century, on the order of three IQ points per decade — is the clearest proof that environments move population means even as heritability of the remaining variance stays high (Flynn, 1987).

Molecular genetics has begun to identify the specific variants. Because intelligence is polygenic — influenced by thousands of variants each of tiny effect — early candidate-gene studies failed, and progress required very large genome-wide association studies (GWAS). A 2017 meta-analysis of about 78,000 individuals identified new intelligence-associated loci and genes expressed in neural tissue (Sniekers et al., 2017); a 2018 study of nearly 270,000 people raised the count sharply, implicating genes involved in neuronal development and synaptic function (Savage et al., 2018). The polygenic scores derived from this work now predict a few percent of the variance in independent samples — far below heritability, a gap (the missing heritability) that shrinks as sample sizes grow (Deary, Cox, & Hill, 2022).

The third demonstration uses the classic twin design to show how the difference between identical- and fraternal-twin correlations yields a heritability estimate.

Worked Example

Falconer's formula estimates the heritability of a trait from twin correlations: h² = 2 × (r_MZ − r_DZ), where r_MZ is the correlation between identical (monozygotic) twins, who share essentially all their genes, and r_DZ is the correlation between fraternal (dizygotic) twins, who share on average half of their segregating genes.

Suppose a study of adult intelligence finds r_MZ = 0.86 and r_DZ = 0.60. The genetic contribution is estimated by doubling the gap: h² = 2 × (0.86 − 0.60) = 2 × 0.26 = 0.52. Just over half the variance in this sample is attributable to genetic differences.

The shared-environment component follows from the same two numbers: c² = 2 × r_DZ − r_MZ = (2 × 0.60) − 0.86 = 1.20 − 0.86 = 0.34, and the non-shared-environment-plus-error term is the remainder, e² = 1 − h² − c² = 1 − 0.52 − 0.34 = 0.14. Applying the same arithmetic to childhood data (say r_MZ = 0.74, r_DZ = 0.54) gives h² = 2 × (0.74 − 0.54) = 0.40 — lower than the adult estimate, the numerical signature of the Wilson effect (Plomin & Deary, 2015). The third demonstration lets a reader vary the two correlations and watch the three components recompute.

Discussion

The scientific case for a general factor of intelligence is, as an empirical matter, settled: the positive manifold is real, g is reliably extracted, and its predictive validity for education, work, health, and longevity is large and replicable (Deary, 2012; Gottfredson, 1997). What remains genuinely open is interpretation. Whether g reflects a single underlying resource — neural efficiency, processing speed, working-memory capacity — or emerges from mutually beneficial interactions among initially independent cognitive processes (the mutualism hypothesis) is not resolved by the factor structure itself, because rival process models can produce the same positive manifold.

The construct also carries a social history that the science must be reported alongside. Intelligence testing has been misused to justify discrimination, and claims about group differences remain scientifically fraught and frequently overstated relative to the evidence (Nisbett et al., 2012). The mainstream position distinguishes sharply between the well-established facts about individual differences and the far weaker, often confounded evidence about the causes of average differences between groups — a distinction that heritability-within-groups does nothing to bridge, since a trait can be highly heritable within each of two groups while the difference between them is entirely environmental.

Current Directions

The field's most active front is molecular. Genome-wide association studies have grown from tens of thousands to hundreds of thousands and now millions of participants, steadily increasing the number of replicated intelligence-associated loci and the variance captured by polygenic scores (Savage et al., 2018). The current agenda is less about discovering that intelligence is polygenic — that is established — than about turning association into mechanism: identifying which genes act in which cell types and developmental windows, and how they shape the neural circuits that support reasoning (Deary, Cox, & Hill, 2022). Genes implicated to date converge on neuronal development, synaptic signalling, and neurogenesis rather than on any single pathway.

A second strand couples this molecular work to imaging. Large biobank samples now permit genetically informed neuroimaging, testing whether the same variants that predict test scores also predict the structural and functional network properties the P-FIT model highlights (Jung & Haier, 2007; Deary, Cox, & Hill, 2022). The unifying goal is a causal chain running from specific genetic variants, through brain development and network efficiency, to the psychometric g measured a century after Spearman first named it — with the honest caveat that polygenic scores still explain far less variance than twin-based heritability implies, and the individual-prediction value of those scores remains modest.

Common Misconceptions

A high heritability means intelligence is fixed and cannot be changed.

Heritability describes the share of variance between people attributable to genetic differences in a given population and environment; it is not a statement about an individual and does not cap malleability. The Flynn effect — a rise of roughly three IQ points per decade through the twentieth century — shows environments shifting population means substantially even where heritability of the remaining variance is high (Flynn, 1987).

IQ tests measure only a narrow, culture-bound kind of school knowledge.

The general factor emerges from any sufficiently broad battery, including non-verbal reasoning tests designed to minimise acquired knowledge, and it predicts outcomes well beyond school — job performance, health, and longevity (Deary, 2012). The belief persists because early verbal tests were culturally loaded; the construct is not.

There are many wholly independent intelligences, so a single score is meaningless.

Proposed separate faculties, where they can be measured with reliable tests, correlate positively rather than independently — the positive manifold reasserts itself, and a general factor re-emerges (Neisser et al., 1996). Distinct broad abilities are real and useful, but they sit beneath g in a hierarchy, not beside it as equals.

Glossary

Behavioral genetics.

The study of genetic and environmental contributions to individual differences in behaviour, using twin, adoption, and molecular designs.

Crystallized intelligence (Gc).

Accumulated knowledge and verbal skill acquired through experience and education; maintained or growing into late life.

Deviation IQ.

A score expressing a person's standing against an age-matched norm on a normal distribution with mean 100 and standard deviation 15.

Fluid intelligence (Gf).

The capacity to reason and solve novel problems independently of learned knowledge; peaks in early adulthood.

Flynn effect.

The sustained rise in raw intelligence-test scores across the twentieth century, roughly three IQ points per decade.

General factor (g).

The shared variance extracted from a broad battery of cognitive tests; the statistical expression of the positive manifold.

Gene–environment correlation.

The tendency of people to select, evoke, and create environments that match their genetic propensities, amplifying heritable differences over development.

Heritability.

The proportion of variance in a trait, within a population and environment, attributable to genetic differences among individuals; a population statistic, not an individual one.

Intelligence test.

A standardized battery of cognitive tasks whose subtests combine into an estimate of general ability.

Parieto-Frontal Integration Theory (P-FIT).

A neuroscientific model attributing intelligence to the efficiency of information flow between parietal and prefrontal cortex.

Polygenic score.

A weighted sum of an individual's trait-associated genetic variants, used to predict a share of the variance in a trait.

Positive manifold.

The empirical finding that scores on any broad set of cognitive tests correlate positively with one another.

Psychometrics.

The theory and technique of psychological measurement, including the reliability and validity of intelligence tests.

Three-stratum model.

The Cattell–Horn–Carroll hierarchy placing g above roughly a dozen broad abilities above dozens of narrow skills.

Wilson effect.

The increase in the heritability of intelligence across development, from about 20% in early childhood to 60% or more in adulthood.

Key Researchers

John B. Carroll (1916–2003). Psychologist at the University of North Carolina at Chapel Hill; his re-analysis of hundreds of ability datasets produced the three-stratum model that anchors the modern CHC framework. Wikipedia - Wikidata

Raymond B. Cattell (1905–1998). Psychologist at the University of Illinois; introduced the fluid–crystallized distinction that underlies contemporary structural models of intelligence. Wikipedia - Wikidata

Ian J. Deary (b. 1954). Professor of differential psychology at the University of Edinburgh; a leader in cognitive epidemiology and the genetics and lifecourse of intelligence differences. Faculty Page - ORCID - Google Scholar - Wikipedia

Howard Gardner (b. 1943). Research professor at the Harvard Graduate School of Education; proposed the theory of multiple intelligences. Faculty Page - Google Scholar - Wikipedia

Richard J. Haier (b. 1946). Professor emeritus at the University of California, Irvine; a founder of the neuroscience of intelligence and co-author of the P-FIT model. Faculty Page - Google Scholar - Wikipedia

Robert Plomin (b. 1948). Professor of behavioural genetics at King's College London; a leading figure in twin studies and the molecular genetics of intelligence. Faculty Page - ORCID - Google Scholar - Wikipedia

Charles Spearman (1863–1945). Psychologist at University College London; discovered the positive manifold and proposed the general factor g in 1904, founding the psychometric study of intelligence. Wikipedia - Wikidata

Robert J. Sternberg (b. 1949). Professor of human development at Cornell University; proposed the triarchic theory and the concept of successful intelligence. Faculty Page - ORCID - Google Scholar - Wikipedia

Frequently Asked Questions

What is intelligence in psychology?
Intelligence is a general mental capacity for reasoning, problem solving, abstract thinking, and learning from experience, inferred from the positive correlation among performances on diverse cognitive tests (Gottfredson, 1997).

What does the general factor g mean?
The general factor g is the shared variance extracted by factor analysis from a broad battery of cognitive tests; it expresses the positive manifold and predicts educational, occupational, and health outcomes (Deary, 2012).

What is the difference between fluid and crystallized intelligence?
Fluid intelligence is the ability to reason and solve novel problems without relying on learned knowledge, while crystallized intelligence is the accumulated knowledge and verbal skill acquired through experience; the two dissociate across the life span (Cattell, 1963).

How is IQ scored?
A modern IQ is a deviation score that ranks a person against an age-matched norm and rescales performance to a normal distribution with a mean of 100 and a standard deviation of 15, so a score of 130 falls at about the 98th percentile (Deary, 2012).

Is intelligence inherited?
Individual differences in intelligence are substantially heritable, around 50% averaged across the life span, and heritability rises from roughly 20% in early childhood to 60% or more in adulthood (Plomin & Deary, 2015).

Why do IQ scores rise over generations?
The Flynn effect describes a sustained rise in raw test scores of about three IQ points per decade through the twentieth century, showing that environmental change can shift population means even when heritability is high (Flynn, 1987).

Where in the brain is intelligence?
Intelligence is not localized to one region; demanding tasks recruit a distributed parieto-frontal network, and the P-FIT model attributes individual differences to the efficiency of information flow within it (Jung & Haier, 2007).

Can genetics predict a person's intelligence?
Genome-wide studies have identified many intelligence-associated variants, but polygenic scores predict only a few percent of the variance in independent samples, far below twin-based heritability, so they cannot forecast an individual's ability with useful accuracy (Savage et al., 2018).

References