Are IQ Tests Accurate? What They Measure Well, and What They Don't

Yes, with conditions. A professionally administered IQ test such as the WAIS-5 or WISC-V gives highly consistent results, with technical manuals reporting reliability above .90 for the full-scale score, and its scores predict school achievement well and job performance more modestly. But even a good score is a range several points wide, and it captures only part of what people mean by intelligence. Online IQ tests vary enormously, and most cannot produce a real IQ at all because they have never been normed on a representative sample.

"Accurate" means two different things

Psychologists split accuracy into two questions:

• Reliability: does the test give consistent results? If you took it twice, or took two halves of it, would the scores agree?
• Validity: does the score mean what it claims to mean? Does it relate to the abilities and real-world outcomes it is supposed to reflect?

A bathroom scale that always reads five pounds light is reliable but not valid. A test can be reliable without being valid, but it cannot be valid without being reasonably reliable. The professional Standards for Educational and Psychological Testing expect publishers to document both.

How reliable clinical IQ tests are

Full-scale scores on the major individually administered batteries are about as consistent as psychological measurements get. Their technical manuals report internal-consistency coefficients in the high .90s for the full-scale IQ (the WAIS-IV manual, for example, reports .98) and test-retest correlations over a few weeks that are somewhat lower but still above .90. Individual subtests are less reliable, because each contains fewer items.

Translated into IQ points, a reliability of .97 implies a standard error of measurement of about 2.6 points, so the 95% confidence interval around a score is roughly ±5. If your full-scale IQ is reported as 112, your true score most likely lies between about 107 and 117. That is precise by psychological standards and still too wide to treat single points as meaningful: a 128 and a 131 are not reliably different results. More on reading those ranges is in interpreting your IQ results.

What IQ scores predict

Validity is judged by what scores relate to. The evidence is strongest for education:

• The 1996 American Psychological Association task force on intelligence (Neisser and colleagues) put the correlation between IQ and school grades at about .50.
• In a study of more than 70,000 English children, Deary and colleagues (2007) found a correlation of about .81 between a general ability factor measured at age 11 and a factor based on national exam results at age 16.
• Strenze's 2007 meta-analysis of longitudinal studies found that IQ measured earlier in life predicted later educational attainment (about .56), occupational level (about .43) and income (about .20).

For job performance, the estimates have been revised downward. Schmidt and Hunter's influential 1998 review put the validity of general mental ability tests at about .51. Sackett and colleagues (2022) argued that this figure relied on over-corrections for range restriction and re-estimated it at about .31, still a useful predictor but ranked below structured interviews in their analysis. The debate continues; see cognitive tests in hiring.

A correlation of .50 is large for social science, yet it means most of the variation in grades between people comes from other things: effort, teaching, health, home circumstances, interest. IQ predicts averages across groups far better than it predicts any one person's life.

Where IQ tests fall short

• They sample intelligence rather than survey it. Tests measure reasoning, knowledge, memory and speed on structured tasks. Creativity, social judgment, practical expertise and persistence are largely outside their reach, a gap discussed in types of intelligence.
• Test-day factors matter. Sleep, illness, anxiety, medication and motivation can all move a score.
• Language and familiarity matter. Verbal and knowledge subtests assume the language and schooling of the norming sample. Someone tested in a second language, or with little formal schooling, may score below their reasoning ability.
• The extremes are less precise. Only about 0.13% of a normal distribution lies above 145, so a norming sample of a few thousand people contains just a handful of such scores. Very high and very low IQs are estimated from thin data, and many tests run out of harder or easier items.
• Norms age. Average performance rose through much of the 20th century (the Flynn effect), so a test with old norms can inflate scores by several points.

Are online IQ tests accurate?

An online test can be reasonably consistent, since well-written items behave much the same on a screen as on paper. What it almost never has is the second ingredient of an IQ score: a representative norm group. Without one, the test can tell you how you compare with the people who happened to take it, not with the population.

What a free online test, including ours, can honestly offer is a rough sense of how you performed on a set of reasoning tasks compared with other people who took the same test. Ours compares you with its own respondents and shows an estimated IQ converted from that position, labelled as an estimate, with a range of ±12 percentile points on the short form and ±7 on the full form, and flags answers given implausibly fast as low-confidence. The sample is still small, so percentiles will move as it grows. It cannot diagnose anything, does not replace a clinical assessment and should not be used for hiring or admissions. The methodology page explains how it works.

How to check whether a particular test is accurate

You do not need access to a technical manual to ask the right questions. For any test, clinical or online, look for:

• A described norm group: who was tested, how many people, when, and whether they were chosen to represent a population or simply volunteered.
• Reported reliability: a coefficient for the overall score, ideally above .90 if the result will be used for anything important.
• A margin of error: a confidence interval or range around the result, not a bare number.
• A stated purpose: what the score is meant for, and what it should not be used for.

A test that cannot answer these questions may still be fun, but its number should not change any decision you make.

Common IQ myths

• "Your IQ is fixed for life." Scores are fairly stable in adulthood, but not fixed. A meta-analysis by Ritchie and Tucker-Drob (2018) estimated that each additional year of education raises IQ by roughly 1 to 5 points. See can IQ change over time.
• "Brain-training games raise your IQ." A major review by Simons and colleagues (2016) found that people improve on the trained games, with little evidence that the gains transfer to broader cognitive abilities or everyday performance.
• "IQ tests only measure how good you are at IQ tests." If that were true, scores would not predict school results years later. They do, as the correlations above show. They also correlate strongly with one another across very different kinds of tasks, the pattern behind the general factor described in the g factor explained.
• "An online test said 145, so I'm a genius." On a properly normed test, 145 or above occurs in roughly 1 in 740 people. If a site hands out scores like that often, it is flattering its users, not measuring them.
• "Einstein's IQ was 160." There is no record that Einstein ever took an IQ test. Figures like this are guesses that have been repeated until they sound like facts.
• "IQ tests are pure cultural bias, so they mean nothing." Language and familiarity with test content clearly matter. But the APA task force concluded that, in the US, the major tests predicted school performance about as well for the groups it examined. Fair prediction is a narrow technical point, not proof that every test-taker is equally familiar with the content.
• "A high IQ guarantees success." The correlation with income in Strenze's meta-analysis was about .20, which leaves enormous room for everything else.