General intelligence, or g, is the common statistical dimension in cognitive performance: the tendency for people who perform relatively well on one kind of cognitive test to perform relatively well on many others.
That sentence is precise, but it still leaves an obvious question: how can several very different tests turn into one thing?
The answer begins with a pattern found repeatedly in data. Give a large group of people a varied set of cognitive tests—for example, tests of abstract reasoning, vocabulary, memory, spatial reasoning and processing speed. These tests are not interchangeable. A vocabulary task asks something different from a spatial task. Yet the scores usually correlate positively. People who perform above average on one tend, on average, to perform above average on others as well.
Psychologists call this pattern the positive manifold. g is a statistical way of representing the part of performance that those tests share.
A worked example: where does g come from?
Imagine 1,000 adults complete five fictional cognitive tasks. Their scores are standardized so that higher numbers mean better performance relative to the group.
| Person | Reasoning | Vocabulary | Memory | Spatial | Processing speed |
|---|---|---|---|---|---|
| A | High | High | High | High | High |
| B | High | Average | High | Average | High |
| C | Average | Average | Average | Average | Average |
| D | Low | Low | Average | Low | Low |
No real dataset looks this neat, but across a large sample a general tendency like this usually appears. The tests might correlate approximately like this:
| Reasoning | Memory | Spatial | Vocabulary | |
|---|---|---|---|---|
| Reasoning | 1.00 | .55 | .62 | .45 |
| Memory | .55 | 1.00 | .40 | .42 |
| Spatial | .62 | .40 | 1.00 | .35 |
| Vocabulary | .45 | .42 | .35 | 1.00 |
The exact numbers above are illustrative, not empirical estimates. What matters is the pattern: the off-diagonal correlations are mostly positive.
A statistical method such as factor analysis asks whether those correlations can be represented efficiently by a smaller number of latent dimensions. In a simple one-factor model, each test receives a factor loading describing how strongly its scores relate to the common factor. A fictional result might look like this:
| Test | Illustrative loading on g |
|---|---|
| Abstract reasoning | .80 |
| Spatial reasoning | .72 |
| Memory | .65 |
| Vocabulary | .61 |
| Processing speed | .45 |
A loading is not the percentage of intelligence contained in a test. It is a statistical relationship between performance on that test and the latent factor in that model.
The important point is that g is derived from the pattern across many people and many tests. Researchers do not look at one person’s results and decide that a hidden substance called g is present. First they estimate the common dimension from the covariance structure of a sample. They can then estimate where individuals stand on that dimension.
Is g an average of all the tests?
Not exactly. A simple average can sometimes approximate general performance, but factor models do something more specific.
An unweighted average gives every test equal influence. A factor model estimates how strongly each observed measure relates to the common dimension and distinguishes variance shared among measures from variance unique to individual measures. That unique component can include both specific influences and measurement error; ordinary factor analysis does not automatically separate those two sources.
This is why a reasoning test might contribute more information about g than another test in a particular battery. It is not because reasoning has been declared “more intelligent.” It is because, in that dataset and model, reasoning performance may correlate more strongly with the common variance shared across the battery.
Different batteries can therefore yield somewhat different g estimates because they sample different tasks. A broad battery is generally preferable when the goal is to estimate general cognitive ability, because it gives the model more diverse information.
What number does g have?
g does not have an intrinsic 1–100 scale.
In research, an individual’s estimated factor score may be expressed in standard-deviation units. For example:
- 0 might represent the sample mean;
- +1.0 means about one standard deviation above that mean;
- −1.0 means about one standard deviation below it.
A researcher could transform those values onto another scale if useful. The scale itself is a convention. What matters is a person’s relative position on the latent dimension estimated by the model.
This is where g and IQ are easy to confuse.
g and IQ: what is the practical difference?
g is a general latent dimension estimated from the covariance among diverse cognitive measures. IQ is a standardized score produced by a particular intelligence test.
Put practically: g is a statistical construct researchers estimate from patterns of covariance; an IQ score is a measurement result from a standardized instrument used to estimate broad cognitive ability.
Suppose a professionally developed intelligence test contains several subtests. A person completes them. Their raw performance is converted using age-appropriate norms, and scores are combined according to the test’s scoring rules into an IQ composite. That composite is usually designed to reflect broad general cognitive ability and will often correlate strongly with psychometric g.
But the IQ score is not mathematically identical to g. It also reflects the particular tasks included, the test’s weighting and scoring system, specific abilities represented in the battery and measurement error.
This means researchers can estimate g from a dataset without calculating anyone’s conventional IQ score. Conversely, a person can receive an IQ score from a standardized test without a researcher calculating a bespoke factor score for that person.
What does a high or low g score actually mean?
A higher estimated g score means that, across the set of cognitive measures represented by the model, a person’s pattern of performance is generally higher relative to the reference sample on the shared dimension.
It does not mean that the person is uniformly excellent at every cognitive task. A person can have relatively high general ability while showing meaningful relative strengths and weaknesses. Likewise, two people with similar general factor scores can differ in their vocabulary, memory, spatial reasoning or processing speed.
This is one reason hierarchical models are useful: they can represent both the common dimension and more specific broad abilities.
Is g a real thing in the brain?
The statistical pattern is real; the claim that g corresponds to one single biological mechanism is not established.
Factor analysis shows that diverse cognitive performances share variance. It does not reveal the biological cause of that covariance. Some theories interpret g as reflecting a broad general capacity. Other theories propose that the positive manifold emerges because many tests recruit overlapping cognitive processes. Mutualism models propose that initially different abilities become correlated through reciprocal developmental interactions. Network accounts describe intelligence as an emergent property of interacting cognitive systems.
These theories disagree about why g appears while accepting much of the same psychometric evidence.
The safest distinction is therefore:
Positive manifold: the observed fact that diverse cognitive tests tend to correlate positively.
g: the latent statistical dimension used to summarize part of that common variation.
Theory of g: an explanation for why that common variation exists.
Confusing those three levels turns a robust empirical result into a stronger causal claim than the data alone justify.
What does g predict?
General cognitive ability is associated with educational learning and achievement and predicts performance in a range of cognitively demanding settings. Its predictive value is one reason g has remained central to intelligence research.
But prediction is probabilistic. A general cognitive score does not determine a person’s education, occupation, income, health or life trajectory. Such outcomes also reflect opportunity, knowledge, personality, motivation, health, social conditions and many other influences.
The key point is that g is not defined by what it predicts. Researchers first observe and model the shared covariance among cognitive performances. They then ask whether that dimension predicts outcomes outside the test battery.
g is not an organ, not a single test and not an IQ number. It is the common statistical dimension that emerges because different cognitive abilities tend to vary together across people.
We can measure that shared pattern with considerable reliability. We can estimate an individual’s relative standing on it. We can show that it predicts meaningful outcomes. What remains scientifically open is the deeper causal question: why do so many different cognitive abilities covary in the first place?
References
- Spearman, C. (1904). General intelligence, objectively determined and measured. American Journal of Psychology, 15, 201–293. DOI
- McMillen, P., & Levin, M. (2024). Collective intelligence: A unifying concept for integrating biology across scales and substrates. Communications Biology, 7, 378. DOI
- Carroll, J. B. (1993). Human Cognitive Abilities. Cambridge University Press.
- van der Maas, H. L. J., et al. (2006). A dynamical model of general intelligence: The positive manifold by mutualism. Psychological Review, 113, 842–861. DOI
- Kovacs, K., & Conway, A. R. A. (2016). Process Overlap Theory. Psychological Inquiry, 27, 151–177.
- Savi, A. O., Marsman, M., van der Maas, H. L. J., & Maris, G. K. J. (2019). The wiring of intelligence. Perspectives on Psychological Science, 14, 1034–1061. —. DOI
