IQ and Crime: Association Without Moral Ranking

0 The Short Answer

Criminology has found a real but modest negative association between measured intelligence and officially recorded offending at the population level. That is the entire finding. It is not a tool for judging a person, and this page exists mainly to explain why, in detail, before it explains anything else.

A population-level correlation describes a pattern averaged across thousands of people. It says nothing reliable about any single person inside that population, because the two groups being compared overlap enormously: most people with lower scores never offend, and a large share of people with higher scores do. A statistic built from thousands of lives cannot be handed back to one life and used as a verdict.

The research also has a structural limitation that most summaries skip entirely: it measures who gets caught, charged and convicted, not who commits an offense. Those are different populations, and the gap between them is shaped by policing intensity, legal resources, and how the justice system treats different people, none of which has anything to do with cognition.

1 What One Major Study Actually Found

The most frequently cited modern data point on this topic comes from Joseph Schwartz, Jukka Savolainen, Mikko Aaltonen and colleagues, published in the journal Intelligence in 2015 (volume 51, pages 109 to 118). The study drew on a total birth cohort of Finnish men born in 1987, tracked through national administrative records rather than a sample the researchers assembled themselves, which is one of its genuine strengths: total-cohort designs avoid the recruitment biases that distort many smaller studies.

Intelligence was measured using three subscales, verbal, mathematical and spatial reasoning, along with a composite score. Offending was measured with nine separate indicators drawn from official criminal justice records, covering different types and levels of severity of recorded offenses. The researchers were specifically interested in the shape of the relationship, not only its existence: whether it was a straight line across the whole range of scores, or whether it curved at the extremes.

What they reported was a pattern that was mostly linear, meaning the association held in a fairly consistent direction across most of the score range, with some evidence of a curvilinear pattern at the very lowest and very highest ends of measured ability. In plain terms, the relationship was not perfectly uniform, and the effect at the extremes looked somewhat different from the effect through the middle of the distribution. That nuance rarely survives when the study gets summarized in a headline, but it matters, because it shows the researchers were reporting complexity rather than flattening it into a single tidy number.

It is worth being explicit about what this single study does and does not establish. It documents an association between intelligence subscales and officially recorded offending in one national cohort, using one country's justice system and one set of administrative definitions. It does not establish that intelligence causes offending, it does not generalize automatically to every country or era, and, as later sections explain, it cannot by itself distinguish a true behavioral pattern from an artifact of how offending gets detected and recorded. Readers curious about what these cognitive subscales are actually measuring can see a fuller explanation on our page on what IQ tests measure.

Finland's national administrative data also makes this cohort an unusually clean test case by the standards of this field, since nearly every resident's contact with schooling, the military conscription board and the justice system is captured in linked government registers rather than pieced together from scattered local sources. That completeness reduces some kinds of sampling error that plague smaller, self-selected samples. It does not reduce the two structural issues covered in the next two sections, detection bias and socioeconomic confounding, because those issues live inside how records get generated in the first place, not in how thoroughly a country happens to store them afterward.

2 The Detection Bias Problem

Every study built on official crime records, including the Finnish cohort study, is measuring a specific and narrower thing than "who committed a crime." It is measuring who was detected, investigated, charged and convicted. Those steps are each their own filter, and each filter can be influenced by factors that have nothing to do with whether an offense actually occurred. Psychiatric registers carry a near identical filter, which is why the mental health literature puts detection and diagnosis bias near the front: more educated, articulate, health aware people are likelier to recognize symptoms, seek assessment and receive a formal label.

Criminologists have known this for decades. Michael Hindelang, Travis Hirschi and Joseph Weis published a landmark comparison of self-reported offending against official records in the American Sociological Review in 1979, using interview and questionnaire data collected in Seattle. Their central finding was that the two measurement approaches broadly agreed with each other at a general level, with correlations between self-reported contact with police and official records typically falling between 0.70 and 0.83. That agreement is reassuring for some purposes, but it also confirms the underlying point: self-report and official-record studies are measuring related but distinct things, and any study using only one of them inherits that method's specific blind spots.

Official records are shaped by where police patrol, which behaviors draw enforcement attention, how much legal representation a defendant can access, and how judges and prosecutors exercise discretion at every stage of a case. None of those factors are properties of the person being processed. They are properties of the system processing them. A person can behave identically to another person and end up with a different official outcome purely because of where they lived, what resources they had, or which stage of the system happened to notice them.

This is exactly why the detection bias problem belongs at the front of any serious discussion of this topic rather than as a footnote at the end. Any correlation calculated from official records is, at least in part, a correlation with how the justice system operates, layered on top of whatever underlying behavioral pattern might also be present. The two layers cannot be cleanly separated after the fact, and no cohort study, however well designed, can fully undo that limitation once the data has already passed through a system that treats people unevenly.

3 The Socioeconomic Confound

A confound is a third factor that influences both variables researchers are trying to relate, making the two look connected to each other when much of the real action is happening elsewhere. In this literature, socioeconomic circumstances are the confound that shows up most consistently, and it is not a subtle one.

Childhood poverty, neighborhood disadvantage, school quality and access to stable resources all independently predict performance on standardized cognitive tests. The same conditions independently predict contact with the criminal justice system, through mechanisms that have nothing to do with cognition: fewer economic alternatives, more heavily policed neighborhoods, less access to legal defense, and greater exposure to the kinds of stress and instability that push life trajectories in worse directions. When two outcomes share so many upstream causes, they will correlate with each other even if neither one causes the other at all.

Long-running longitudinal projects that track many risk factors at once, such as David Farrington's Cambridge Study in Delinquent Development, which has followed a cohort of London men since the early 1960s, illustrate why this matters in practice. These studies consistently find that measured intelligence travels alongside a cluster of other circumstances, family income, parental supervision, housing stability, school engagement, rather than appearing on its own. Pulling one variable out of that cluster and studying it in isolation risks assigning it explanatory weight that actually belongs to the conditions surrounding it.

• Shared upstream causes. Poverty and disadvantage affect test performance and justice system contact through separate, well documented pathways, which is enough on its own to produce a correlation between the two outcomes.
• Cluster, not isolate. Intelligence rarely moves independently of income, schooling and neighborhood in longitudinal data, which means single-variable explanations tend to overstate what any one factor is doing.
• Direction is not settled. Disadvantage can suppress measured performance on a given test day just as plausibly as any underlying ability difference can, and cohort studies built from a single testing occasion cannot fully rule that out.

None of this means socioeconomic status fully explains the association away, and no honest reading of the literature claims that it does. It means the association cannot be interpreted as a clean, isolated signal about cognition until the shared causes running underneath both variables are accounted for, and most public discussion of this topic skips that step entirely.

4 Why a Population Correlation Cannot Predict a Person

This is the statistical distinction that the rest of this page depends on, so it deserves a plain explanation rather than a formula. A correlation is a summary of a pattern across a large group. It tells you that, on average, as one variable moves, the other tends to move somewhat in a particular direction. It does not tell you what any specific member of that group will do, because a summary statistic, by design, throws away the information that would let you say that. The point holds just as firmly for associations nobody finds threatening: the childhood IQ and mortality research reports something on the order of a 20 to 24 percent lower risk of death per standard deviation of childhood score, built from whole-country school records, and still forecasts nothing about how long any one person will live.

Picture two overlapping bell curves, one for people who eventually have an official offense on record and one for people who do not. A modest negative correlation shifts those two curves slightly apart on the intelligence scale. It does not separate them into two distinct groups. The overwhelming majority of each curve sits directly on top of the other curve. Someone near the middle of either distribution could plausibly belong to either group, and even someone near one tail still has meaningful overlap with the other.

This is the same reasoning that applies to every modest population correlate of behavior, including the ones with far more benign reputations. A person's score correlates modestly with job performance across large samples, and yet nobody would tell an individual jobseeker their career outcome from a single test result, because the same overlap problem applies there too. A correlation is evidence about a population's tendencies. A prediction about one person requires far more information than a population tendency can supply, and pretending otherwise is a category error, not a bolder reading of the same evidence.

Base rates make the error worse in the crime context specifically. Officially recorded serious offending is, for any given birth cohort, a relatively rare outcome. When an outcome is rare, even a real and statistically significant risk factor produces enormous numbers of false positives if you try to use it predictively, because the vast majority of people who share that risk factor still never produce the outcome. Multiplying a modest correlation by a low base rate does not sharpen a prediction. It guarantees that most individual predictions built from it will be wrong.

5 Reading the Numbers: What "Modest" Really Means

Vague words like "modest" and "weak" get thrown around in science reporting without ever being pinned to a number, which lets readers imagine whatever strength of effect fits their prior assumptions. The statistics underneath are not complicated, and seeing them laid out removes most of the guesswork.

A correlation coefficient, written r, ranges from negative one to positive one. Squaring it, r-squared, gives the share of variance in one variable that is statistically associated with the other. Jacob Cohen's widely used benchmarks from his 1988 text on statistical power treat a correlation around 0.10 as small, around 0.30 as moderate, and around 0.50 as large for the kind of individual-difference research this field works with. Most reported correlations between cognitive scores and officially recorded offending fall inside or below that small-to-moderate range, not near the large end.

Set against that scale, a correlation in the small-to-moderate range is not nothing, and researchers are right to keep studying it. But it is nowhere near strong enough to function as a predictive tool for an individual, and treating it as one misreads what the number in front of you actually says. Anyone wanting a fuller sense of how scores are distributed across a population can see our percentile chart for context on how much variation sits inside every point on the scale.

6 What Low Scores Do Not Mean

It is worth stating this directly and without hedging, because indirect language is how misuse creeps into a topic like this one. A below-average score on a cognitive test, on ACIS or on any other instrument, is not evidence that a person has committed, will commit, or is inclined toward criminal behavior. It is not a character assessment. It is not a risk score. It is a measurement of performance on a specific set of tasks, on a specific day, nothing more. The words about a specific day are doing real work there, because performance measured during a depressive episode can sit below a person's usual level once attention, mental speed and executive control are affected, which is one more reason a single low result describes a session rather than a person.

The confusion usually comes from collapsing a population statistic into a personal one, the same error covered in the previous two sections, but it is worth restating in the specific context of how a person might feel reading this page about their own results. Knowing that a modest negative association exists at the population level tells you nothing about where any one person's abilities, choices or future actually sit. The overwhelming majority of people across the entire range of measured intelligence never have contact with the criminal justice system at all, and that fact alone should end any temptation to treat a test score as a warning sign.

There is also a fairness problem with turning this research into personal judgment that goes beyond bad statistics. Doing so effectively punishes people twice for the same disadvantage: first through whatever socioeconomic conditions suppressed a test score or increased justice system contact in the first place, and second through being viewed with suspicion because of a number that reflects those conditions rather than character or intent. Readers who want a grounded sense of what a given score actually represents, and what it does not, can see our explainer on what counts as a good IQ score, which walks through the same interpretive caution that applies here.

7 A History That Demands Extra Caution

This topic carries a documented history of misuse, and understanding that history is part of reading the current research responsibly rather than a digression from it. Italian physician Cesare Lombroso published his theory of the "born criminal" in 1876, arguing that criminality was a hereditary, identifiable trait detectable through physical characteristics. The theory was influential for decades and was eventually abandoned because it did not hold up to scrutiny, but it set a template that later researchers, working with cruder versions of intelligence testing, repeated in a different form.

American psychologist Henry Goddard published "Feeble-Mindedness: Its Causes and Consequences" in 1914, arguing that low measured intelligence was a root cause of criminality, prostitution and other social problems. That argument fed directly into the American eugenics movement of the early twentieth century, which used intelligence testing to justify forced institutionalization and compulsory sterilization of people labeled "feeble-minded," a policy upheld by the United States Supreme Court in the 1927 case Buck v. Bell. None of that history involved rigorous science by modern standards. All of it used the language of measurement to justify decisions that caused enormous, irreversible harm to real people.

Modern researchers working in this area, including the authors of the Finnish cohort study discussed above, operate under very different methodological standards and very different ethical norms than Goddard did, and it would be inaccurate to equate contemporary criminology with that history. But the history is exactly why a modern, careful, appropriately hedged finding still deserves handling with unusual care before it reaches the public. The distance between "a modest statistical association exists in aggregate data" and "this trait explains and predicts criminality" has been crossed before, with consequences that took decades to undo, and it can be crossed again by anyone who strips the caveats out of a legitimate study for a punchier headline.

8 The Crime Official Records Miss

Official crime statistics are also selective about which kinds of offenses they capture well, and that selectivity interacts directly with the detection bias problem described earlier. American sociologist Edwin Sutherland coined the term "white-collar crime" in a 1939 address to the American Sociological Society, later expanded into his 1949 book of the same period, arguing that offenses committed by people in positions of trust, occupational fraud, financial misrepresentation, regulatory violations, were systematically underrepresented in official crime data relative to how much harm they caused.

That observation still holds. Offenses that require access to institutions, financial systems or professional credentials tend to be investigated through slower, less visible channels than street-level offenses, and they are disproportionately committed by people with more education, more resources and more institutional standing, not less. A dataset built from arrest and conviction records will therefore structurally undercount exactly the category of offending most associated with higher socioeconomic status and, on average, higher measured cognitive scores, while overcounting the categories of offending that draw the most direct and visible policing.

This is not a minor technical footnote. It means that any correlation calculated from official offending records is, by construction, comparing intelligence against a partial and unevenly sampled slice of actual offending behavior, not against the full range of unlawful conduct in a society. A more complete accounting of offending, one that fully captured financial and institutional crime alongside street-level offending, would not necessarily produce the same pattern, and no study to date has been able to fully close that gap.

9 What Might Plausibly Connect Cognition and Behavior

It is possible to take the research seriously without pretending the association is meaningless, and doing so means being precise about what remains a hypothesis rather than an established mechanism. Developmental psychologist Terrie Moffitt proposed an influential framework in a 1993 paper in Psychological Review, distinguishing between offending that is limited to adolescence, which she argued is common and largely a byproduct of normal teenage social dynamics, and a smaller, more persistent pattern that she linked theoretically to early neuropsychological difficulties interacting cumulatively with a difficult environment over the course of development.

Within that kind of framework, researchers have theorized that specific cognitive processes, rather than a single global intelligence number, might plausibly relate to behavior: weaker executive function and self-regulation could make it harder to inhibit impulsive responses in high-stakes moments, and weaker verbal reasoning could make it harder to navigate school, employment systems and eventually legal processes in ways that compound disadvantage over time. These are reasonable theoretical proposals actively studied in the field, not settled causal findings, and framing them as anything more definite than that would misrepresent where the science currently stands.

It is also worth separating the distinct cognitive domains involved rather than treating "intelligence" as one undifferentiated trait, since the domains behave differently in this literature and in general. Fluid reasoning, the capacity to solve novel problems under time pressure, and crystallized, learned verbal knowledge are related but separable constructs, a distinction covered in more depth on our page about fluid versus crystallized intelligence. The Finnish cohort study's decision to examine verbal, mathematical and spatial subscales separately, rather than collapsing everything into one score, reflects the same recognition that a composite number can obscure more than it reveals.

None of these plausible mechanisms have been demonstrated to operate independently of the socioeconomic and detection factors covered earlier in this page. They remain candidate explanations sitting alongside, not replacing, the confounds and measurement issues already discussed, and any responsible reading of this literature holds all of those possibilities in mind at once rather than picking the one that sounds most like a discovery.

10 Risk Assessment Tools and Why Scores Should Stay Out of the Courtroom

This research does not exist in a vacuum. Actuarial risk assessment tools are already used in some parts of the criminal justice system to inform decisions about bail, sentencing and parole, and those tools have faced serious, well documented criticism. A 2016 investigation by journalists at ProPublica examined one widely used commercial risk-scoring tool and found that its error patterns differed across groups of defendants in ways that raised significant fairness concerns, sparking an ongoing debate among researchers, statisticians and legal scholars about whether any statistical tool of that kind can be fair once real-world data, shaped by the same detection and socioeconomic issues covered above, feeds into it. Legal settings and the coverage around them also import an older shorthand, the claim that a defendant has the mental age of a child, and that age equivalent framing carries no diagnostic authority today: intellectual disability is evaluated through intellectual and adaptive functioning judged by qualified professionals, not through a number of years.

A cognitive test score is even less suited to that kind of use than the tools already under scrutiny, because it was never designed or validated for the purpose of predicting individual criminal behavior, and folding it into a bail hearing, a sentencing decision or a hiring background check would compound every statistical problem described on this page: a modest correlation, a confounded relationship, a detection-biased outcome variable, and a rare base rate, all stacked together and then applied to a single human being who deserves to be judged on their own conduct.

ACIS was built as a self-assessment of cognitive ability across separate domains, reported with intervals rather than a single verdict number, and it was never designed, validated or intended for use in any legal, employment screening, or risk-related decision about a person's likelihood of offending. No legitimate cognitive test on the market today was built for that purpose, and any product or service claiming otherwise should be treated with the same skepticism this page applies to every other unverified claim in this field.

11 How to Read Any Study on This Topic Responsibly

The tools used on this page generalize to any future headline claiming a link between intelligence and crime, and it is worth keeping them close at hand, because this is a topic that will keep generating new studies and new misreadings of them. The same rules apply to other charged correlates, such as the religiosity research, where measurement problems sit on both sides of the correlation.

• Check whether the outcome is behavior or detection. Official records measure processing by the justice system, not the underlying rate of offending. A study should say explicitly which one it used, and a fair reading should note the difference out loud.
• Check whether socioeconomic factors were addressed. A study that ignores income, neighborhood and education has not isolated intelligence from the conditions that travel alongside it.
• Check the actual size of the correlation. A vague word like "linked" or "associated" can describe an r of 0.05 or an r of 0.50, and those describe extremely different realities.
• Check whether the claim stays at the population level. Any sentence that converts a group average into a statement about what one person will do has already left the science behind.
• Check for group comparisons dressed up as findings. A responsible study of this topic describes a pattern within a population, not a ranking between populations, and any version that does the latter should be read with active suspicion regardless of its source.

Applied to the research summarized on this page, the honest verdict is narrow: a real, modest, population-level association has been documented in credible data, it is very likely partly a product of detection bias and socioeconomic confounding, and it supports no conclusion whatsoever about any individual person or any group ranking. That is a much smaller claim than the one implied by most headlines on this subject, and it is the only claim the evidence actually supports.

12 Verdict

The research on intelligence and crime is real, it comes from credible cohort data including the Finnish 1987 birth cohort analyzed by Schwartz and colleagues, and it describes a modest negative association at the population level between measured intelligence and officially recorded offending. That is the whole finding, and everything built on top of it, from historical eugenics policy to modern risk-assessment tools, has come from stretching that modest finding far past what it can support.

Detection bias means official records track processing by the justice system as much as they track behavior. Socioeconomic confounding means poverty and disadvantage plausibly drive both variables independently. Overlapping distributions and low base rates mean a population correlation, however real, cannot be turned into a statement about any one person. Put those three limitations together and what remains is a legitimate, narrow, still-developing area of research, not a diagnostic tool, not a ranking system, and not grounds for suspicion about anyone's character based on a test score.

If this page leaves you with one number worth remembering, it should not be a correlation coefficient. It should be this: the overwhelming majority of people across the entire range of measured cognitive ability never have any contact with the criminal justice system, which is the simplest and most reliable fact this whole literature confirms.

If you came to this page curious about your own cognitive profile rather than about crime statistics, that is a separate and much more constructive question, and it is the one ACIS is actually built to answer: a profile across distinct domains, reported with honest uncertainty, not a verdict about who you are.

13 Frequently Asked Questions

14 Sources Behind This Page

The relationship discussed here comes from published research, and the honest reading includes its limits. These are the primary sources behind the numbers on this page.

• Gottfredson, L.S. (1997). Why g matters: the complexity of everyday life. Intelligence, 24(1), 79-132.
• Strenze, T. (2007). Intelligence and socioeconomic success: a meta-analytic review of longitudinal research. Intelligence, 35(5).
• Nisbett, R.E. et al. (2012). Intelligence: new findings and theoretical developments. American Psychologist, 67(2). The broad APA review of what moves measured intelligence and what does not.
• Schmidt, F.L. & Hunter, J.E. (1998). The validity and utility of selection methods in personnel psychology. Psychological Bulletin, 124(2). The meta-analysis behind general ability as the strongest single predictor of job performance.
• Deary, I.J., Strand, S., Smith, P. & Fernandes, C. (2007). Intelligence and educational achievement. Intelligence, 35(1). A five year cohort of over 70,000 English schoolchildren.
• Kuncel, N.R., Hezlett, S.A. & Ones, D.S. (2004). Academic performance, career potential, creativity, and job performance: can one construct predict them all? Journal of Personality and Social Psychology, 86(1).
• Whalley, L.J. & Deary, I.J. (2001). Longitudinal cohort study of childhood IQ and survival up to age 76. BMJ, 322(7280). Open access.
• Batty, G.D. et al. (2007). IQ and mortality: a population based cohort study. International Journal of Epidemiology, 36(3), 538-553.
• Zagorsky, J.L. (2007). Do you have to be smart to be rich? The impact of IQ on wealth, income and financial distress. Intelligence, 35(5).
• Judge, T.A., Colbert, A.E. & Ilies, R. (2004). Intelligence and leadership: a quantitative review and test of theoretical propositions. Journal of Applied Psychology, 89(3).