The g Factor: What General Intelligence Actually Is — And What It Isn't

Here is the finding that started it all, and it is genuinely strange. Take a group of people. Give them a vocabulary test, a mental rotation task, an arithmetic problem set, and an abstract pattern puzzle. These tasks have almost nothing in common on the surface. Yet the people who do well on one tend to do well on all the others. Every correlation comes out positive. Always.

That pattern is called the positive manifold, and it has been described as arguably the most replicated result in all of psychology. The g factor is the name for whatever it is that the tests share. Understanding g is the difference between knowing what an IQ score actually is and merely knowing your number.

But — and most explainers stop right before this part — the fact that the manifold exists does not automatically mean a single general ability causes it. That gap is where a century of argument lives, and it's where this article spends its time.

Want a real estimate rather than a theory lesson? Our free IQ test uses both verbal and non-verbal items — deliberately, because as you'll see below, a mixed battery is the only way to get near g at all.

Spearman and the positive manifold

In 1904, Charles Spearman published a paper with an awkward title and an enormous legacy: "General Intelligence," Objectively Determined and Measured. His data was modest by any modern standard — a small group of village schoolchildren, tested on sensory discrimination and rated on school subjects including Classics, French, English, mathematics, and pitch discrimination.

The inter-test correlations landed roughly between +0.25 and +0.55. What mattered wasn't the size. It was the sign. Every single one was positive. If the mind were a collection of genuinely independent modules — a language faculty here, a number sense there — those correlations should have scattered around zero. They didn't.

Spearman drew the obvious inference: a common cause. He proposed a two-factor theory — performance on any mental task equals a general factor g, shared by all tasks, plus a specific factor s, unique to that task. Two tests correlate to the extent that both draw on g. To extract that shared component from a correlation matrix, he had to invent a statistical technique along the way: factor analysis.

One consequence Spearman drew has held up remarkably well — the indifference of the indicator. Because all cognitive tests tap g to some degree, it matters surprisingly little which tests you use, as long as the battery is broad. Estimate g from vocabulary and arithmetic, or from matrices and mental rotation, and you land in a similar place.

How much does g actually explain?

This is where careful sourcing matters, because the number gets inflated constantly. The authoritative estimate comes from John Carroll's 1993 re-analysis of more than 460 datasets: across a diverse cognitive battery, g accounts for roughly 40–50% of total variance. Jensen's (1998) figure for the first principal component runs in a similar band, often quoted as 40–60% depending on the battery.

Read that in both directions, because both directions are true:

• g is substantial. No other single dimension in differential psychology explains anything close to half the variance in a domain that broad.
• g is not most of you. Half or more of the variance in test scores belongs to something else — broad abilities, narrow skills, test-specific quirks, and plain measurement error. Anyone claiming an IQ score captures "your intelligence" is overselling by at least a factor of two.

This is also why arguments about emotional intelligence and multiple intelligences aren't automatically anti-scientific. The g theorist's claim was never that g is everything. It's that g is the largest single common thread.

Where g sits: the CHC hierarchy

Modern intelligence research doesn't treat g as a rival to specific abilities. It stacks them. The dominant framework — the Cattell–Horn–Carroll (CHC) model — is a three-level hierarchy, and nearly every major IQ test built in the last thirty years is organized around it.

The hyphenated name marks a real convergence. Raymond Cattell's fluid–crystallized distinction, expanded by John Horn into nine or ten broad factors, arrived at nearly the same middle layer that Carroll's independent 460-dataset survey produced. Two research programmes, different methods, converging structure. Kevin McGrew proposed merging them in the late 1990s, and the synthesis stuck (McGrew, 2009).

The practical upshot: a single IQ number compresses this entire hierarchy into one digit. That's efficient, and it throws away nearly all the interesting information — which is precisely why domain-level profiles exist.

What g predicts — and the correction nobody quotes

The strongest practical argument for g has always been criterion validity: it predicts things people care about. For decades, the canonical citation was Schmidt and Hunter's 1998 meta-analysis, summarizing 85 years of personnel-selection research, which reported an operational validity of r = .51 for general mental ability predicting job performance — rising to about .58 in high-complexity jobs.

That number is quoted everywhere. What's quoted far less is that it has been substantially revised.

Sackett and colleagues (2022) re-examined the statistical corrections underlying those estimates — specifically, how aggressively to correct for range restriction and criterion unreliability — and concluded the older figures were inflated. Their revised estimate for cognitive ability was r = .31, which moved it out of first place in the validity hierarchy; structured interviews came out on top. A follow-up analysis restricted to studies from 2000–2021 suggested a figure closer to .22. Sackett has called this work a course correction for the field, and it is genuinely contested — critics argue that declining to correct for range restriction in concurrent studies is itself an extreme methodological position.

Two honest conclusions follow. First, g does predict real outcomes — job performance, educational attainment, occupational level, income, and even health and longevity, a relationship Deary and colleagues have documented extensively. Second, the effect sizes are smaller than a generation of textbooks claimed, and the difference between .51 and .31 is not a rounding error — it's the difference between "dominant predictor" and "one useful predictor among several."

The serious challenges to g

Here's the part that separates a real explainer from a summary. The positive manifold is not in dispute. What's in dispute is the inference from "all tests correlate" to "one general ability causes it." Several models reproduce the manifold with no g at all:

Sampling theory (Thomson, 1916)

Godfrey Thomson showed, almost immediately after Spearman, that if each mental test samples randomly from a huge pool of independent neural "bonds," tests will overlap in what they sample — producing a positive manifold that looks exactly like g, with no general ability anywhere in the model. Bartholomew and colleagues revisited this argument in 2009. It has never been decisively refuted; it's mathematically equivalent to g on the data Spearman had.

Mutualism (van der Maas et al., 2006)

This dynamical model proposes that cognitive abilities start out uncorrelated and become correlated during development, because they mutually support one another — better memory helps you build vocabulary, vocabulary helps reasoning, reasoning helps memory strategies. Run the model forward and a positive manifold emerges spontaneously. In the authors' words, a single underlying g factor plays no role in the model. On this account g is an epiphenomenon — a real pattern, but a consequence rather than a cause.

It has taken fire. Gignac (2014) tested a mutualism prediction — that g should be weak or absent in very young children and strengthen with age — and found the increase in g's strength from age 2.5 to 10 was weak in magnitude, which doesn't sit comfortably with the model. The debate remains open, in part because mutualism is difficult to falsify with standard confirmatory factor analysis.

Process overlap theory (Kovacs & Conway, 2016)

The most influential recent alternative. It argues that many cognitive tasks tap overlapping executive processes — domain-general attentional control is required by nearly every difficult item — so tests correlate because they share processes, not because they share a general ability. A key implication: IQ is better understood as an emergent formative construct than as a reflective latent trait. In plainer terms, g would be a summary of overlapping demands, not an underlying quantity you possess.

And separately, the reification critique — most famously pressed by Stephen Jay Gould — holds that extracting a mathematical factor and then treating it as a thing in the head is a category error. Factor analysis finds structure; it doesn't tell you what caused the structure. That objection is philosophically sharper than its reputation among g's defenders suggests, even if Gould's broader historical claims have drawn heavy criticism.

Is g a thing in the brain?

The evidence here is real but more modest than headlines imply. General cognitive ability is substantially heritable and massively polygenic — genome-wide association studies consistently find that it arises from a very large number of variants, each with a tiny effect, rather than a handful of "intelligence genes." Large GWAS efforts have implicated processes like neurogenesis and myelination, and g correlates with neural measures including brain volume and processing efficiency.

But heritability doesn't settle the causal-structure question. A trait can be highly heritable and still be an emergent product of many interacting subsystems — which is exactly what mutualism predicts. And genetic influence on g the score doesn't demonstrate that g is a unitary biological quantity. The Flynn effect makes this concrete: population scores rose ~3 points per decade for a century through purely environmental change, and meta-analytic work suggests those gains were largely not on g — which is itself evidence that observed scores and the latent factor can move independently.

The honest verdict

Three things are true at once, and you have to hold all three:

• The positive manifold is bedrock. Cognitive tests correlate positively, everywhere, always. Any theory of intelligence has to explain this. It is not going away.
• g is a robust statistical regularity with real predictive power — roughly 40–50% of battery variance, and genuine (if smaller than advertised) prediction of educational, occupational, and health outcomes.
• The causal interpretation of g is genuinely unsettled. Whether it's one general ability, overlapping processes, or a developmental epiphenomenon is an open scientific question, not a solved one.

Which means the right way to read your own score sits between two errors. Dismissing IQ as meaningless ignores a century of replicated structure. Treating it as a measurement of your fundamental worth ignores that even its defenders only claim it captures about half the variance in a narrow slice of cognition — and that the number itself carries real measurement error, as we cover in IQ test accuracy. For how raw responses become that number in the first place, see how IQ tests are scored.

Frequently asked questions

Related reading

← Browse all articles on the blog

References

1. Spearman, C. (1904). "General intelligence," objectively determined and measured. American Journal of Psychology, 15(2), 201–292.
2. Thomson, G. H. (1916). A hierarchy without a general factor. British Journal of Psychology, 8(3), 271–281.
3. Horn, J. L., & Cattell, R. B. (1966). Refinement and test of the theory of fluid and crystallized general intelligences. Journal of Educational Psychology, 57(5), 253–270.
4. Carroll, J. B. (1993). Human Cognitive Abilities: A Survey of Factor-Analytic Studies. Cambridge University Press.
5. Carroll, J. B. (2003). The higher-stratum structure of cognitive abilities: Current evidence supports g and about ten broad factors. In H. Nyborg (Ed.), The Scientific Study of General Intelligence (pp. 5–21). Pergamon.
6. Jensen, A. R. (1998). The g Factor: The Science of Mental Ability. Praeger.
7. Schmidt, F. L., & Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology: Practical and theoretical implications of 85 years of research findings. Psychological Bulletin, 124(2), 262–274.
8. van der Maas, H. L. J., Dolan, C. V., Grasman, R. P. P. P., Wicherts, J. M., Huizenga, H. M., & Raijmakers, M. E. J. (2006). A dynamical model of general intelligence: The positive manifold of intelligence by mutualism. Psychological Review, 113(4), 842–861.
9. Bartholomew, D. J., Deary, I. J., & Lawn, M. (2009). A new lease of life for Thomson's bonds model of intelligence. Psychological Review, 116(3), 567–579.
10. McGrew, K. S. (2009). CHC theory and the human cognitive abilities project: Standing on the shoulders of the giants of psychometric intelligence research. Intelligence, 37(1), 1–10.
11. Deary, I. J. (2012). Intelligence. Annual Review of Psychology, 63, 453–482.
12. te Nijenhuis, J., & van der Flier, H. (2013). Is the Flynn effect on g? A meta-analysis. Intelligence, 41(6), 802–807.
13. Gignac, G. E. (2014). Dynamic mutualism versus g factor theory: An empirical test. Intelligence, 42, 89–97.
14. Kovacs, K., & Conway, A. R. A. (2016). Process overlap theory: A unified account of the general factor of intelligence. Psychological Inquiry, 27(3), 151–177.
15. Savage, J. E., et al. (2018). Genome-wide association meta-analysis in 269,867 individuals identifies new genetic and functional links to intelligence. Nature Genetics, 50(7), 912–919.
16. Kovacs, K., & Conway, A. R. A. (2019). What is IQ? Life beyond "general intelligence." Current Directions in Psychological Science, 28(2), 189–194.
17. Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology, 107(11), 2040–2068.
18. Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2024). Revisiting the design of selection systems in light of new findings regarding the validity of widely used predictors. Industrial and Organizational Psychology, 17(1), 1–21.