Humor and measured ability

0 Quick Answer

Being funny does track measured ability, but only when funniness is measured by having people write jokes that strangers then rate, and the association runs through verbal knowledge rather than through reasoning with unfamiliar material. The cleanest decomposition is Christensen, Silvia, Nusbaum, and Beaty, reported in Psychology of Aesthetics, Creativity, and the Arts in 2018 on 270 adults. Humor production ability correlated .49 with crystallized ability, .38 with broad retrieval ability, and .22 with fluid reasoning. A general factor estimated from all three predicted humor production at .51.

That is a substantial relationship by the standards of individual differences research. It is also a relationship with a specific shape that popular coverage flattens. Producing humor is not the same skill as appreciating it, and the evidence for the two is not equally strong. The correlation sits with verbal ability more than with fluid reasoning, and in a bifactor model the fluid effect disappears entirely once general ability is accounted for. And humor production is measured with cartoon caption tasks scored by trained raters, never by asking people whether they are funny, because self-ratings barely track judged performance.

This page is about the research on humor and measured ability. Where humor sits among the behavioral markers people look for is covered separately in the list of signs of high intelligence, which owns that framing. What follows here is the studies, their samples, their effect sizes, and the two places where the popular version of this finding goes wrong.

1 What a Humor Study Actually Measures

Every result on this page rests on the same procedure, and understanding the procedure explains most of what the results can and cannot support. Participants are given a stimulus, usually a cartoon with its caption removed, and told to write something funny. Their attempts are then rated for funniness by judges who do not know who wrote them and who never see the writer's test scores. The rated funniness is the dependent variable. Nobody is ever asked whether they have a good sense of humor.

The details vary and the variation matters. Greengross and Miller used captions for three cartoons rated by six judges, and reported that interrater agreement averaged 0.72 across the three. Christensen and colleagues used three different task types, cartoon captions, joke stems, and satirical definitions, and scored them with a many facet Rasch model that adjusts each person's estimate for how lenient or severe the raters who happened to see it were. Their person reliability estimates were .68 for definitions, .63 for jokes, and .45 for cartoon captions, which is a direct warning that the single most popular task in this literature is also the noisiest one.

Howrigan and MacDonald went further from the cartoon convention, using six character profiles, three sarcastic email responses, and two pages of humorous drawings. Their eleven item composite reached a Cronbach's alpha of .74 with a mean interrater correlation of .42, while the individual items averaged .58 and .28. That gap is the whole reason these studies aggregate across many attempts: one joke tells you almost nothing about a person, and eleven of them tell you something.

The consequence is that a low score on one of these tasks is weak evidence about a person and a group difference across hundreds of them is much stronger evidence about a population. That asymmetry runs through every outcome finding, and the general logic of it is set out on the page on measured ability and life outcomes.

2 Producing Humor and Enjoying It Are Different Findings

The single most common error in coverage of this topic is treating a study of humor production as though it were a study of humor appreciation, or the reverse. They are different tasks with different demands and different evidence behind them, and the strong result belongs to production.

Production means generating something funny that did not exist a moment ago. It requires retrieving relevant material from memory, holding several possible directions in mind, selecting the one with the sharpest incongruity, and expressing it compactly. Appreciation means recognizing that something already constructed is funny. The construction work has been done for you; what remains is detecting the incongruity and resolving it.

Every effect size on this page above .20 comes from a production study. The appreciation literature is smaller, uses ratings of existing material, and reports weaker and less consistent relationships with measured ability, in part because appreciation ratings are compressed. Most people find most professionally written jokes at least somewhat funny, so there is less variance to explain. Production tasks spread people out because most attempts fail.

• Production tasks give a stimulus and require an original response, which is then rated by strangers. This is where the .29 to .51 range of effects comes from.
• Appreciation tasks present finished material and record ratings of funniness, difficulty, or comprehension. This is what the black humor research measured.
• Self-report humor scales ask people about their own humor styles or abilities. These measure beliefs about oneself, which is a different variable again.

Keeping the three apart resolves most of the apparent contradictions in popular articles on this topic. A study finding that people who enjoy dark humor score higher and a study finding that people who write funnier captions score higher are not two measurements of the same thing, and neither one licenses the other's headline. The same distinction between generating and recognizing runs through the adjacent literature on measured ability and creativity, where divergent production and recognition tasks also behave differently.

3 The Study That Set the Reference Point

Greengross and Miller established the association in a sample large enough to split by sex and still have power in each half. Their paper appeared in Intelligence in 2011, volume 39, pages 188 to 192. Four hundred University of New Mexico students took part, 200 men and 200 women. Abstract reasoning was measured with Raven's Advanced Progressive Matrices, verbal ability with the vocabulary subtest of the Multidimensional Aptitude Battery, and humor production with the rated funniness of captions written for three cartoons.

The headline correlations were .38 between vocabulary score and humor ability, and .25 between Raven's score and humor ability, both at p below .001. The authors tested whether the difference between those two was real using a Fisher r to z transformation and reported z = 2.1, significant at p below .05. That single comparison is the origin of the claim that the humor link is verbal, and it is worth noting that the authors ran the test rather than eyeballing the gap.

Split by sex, the pattern held in both halves. For men, humor ability correlated .44 with vocabulary and .27 with Raven's. For women, the same correlations were .31 and .24. The structural equation models produced latent effects of general intelligence on humor ability of .67 for men and .51 for women, figures later described by Christensen and colleagues as among the large effects in this literature.

The sex difference in humor production itself was smaller than the intelligence association. Men averaged 0.09 on the standardized humor measure with a standard deviation of 0.48, women averaged minus 0.09 with a standard deviation of 0.49, giving t = 3.77 and a Cohen's d of 0.38. That is a real difference in the sample and a modest one, with the two distributions overlapping across most of their range.

4 The Study That Took the Association Apart

The most informative paper in this literature is the one that stopped asking whether intelligence predicts humor and started asking which ability does. Christensen, Silvia, Nusbaum, and Beaty ran 270 young adults at the University of North Carolina at Greensboro through three humor tasks and three families of cognitive tasks, then modeled the result inside the CHC framework. Mean age was 19.08 with a standard deviation of 3.14, and 86 percent of the sample was women.

Fluid reasoning was measured with paper folding, series completion, letter sets, and number series. Crystallized ability was measured with two vocabulary tests. The addition that made the paper matter was broad retrieval ability, measured with five fluency tasks in which participants generate as many instances of a category as they can. That factor had never been examined in humor research before.

The latent correlations came out as follows. Crystallized ability correlated .49 with humor production, with a confidence interval from .29 to .70. Broad retrieval correlated .38, interval .22 to .54. Fluid reasoning correlated .22, interval .04 to .39, the only one of the three whose interval came close to zero. A confirmatory model fit the data well, with a chi-square of 112.768 on 72 degrees of freedom, a CFI of .937, an RMSEA of .046, and an SRMR of .050.

That last row is the finding that changes how the whole literature should be read. In the bifactor model, the general factor predicted humor production at .47, broad retrieval kept a specific effect of .20, and fluid reasoning's specific effect collapsed to .09 and lost significance. The authors state the implication directly: the variance in the fluid tasks that predicts humor ability is the variance due to the general factor, not the variance unique to fluid reasoning. Crystallized ability, meanwhile, was absorbed into the general factor entirely, because the general factor explained the vocabulary tasks so thoroughly that no separate crystallized factor remained.

5 Every Study Side by Side

Laid out together, the studies agree on the direction and disagree on the size in ways that track their measures. The pattern across four decades is a correlation somewhere between .22 and .51 depending on which ability is measured, how many humor attempts are aggregated, and whether the estimate is a raw correlation or a latent one corrected for measurement error.

Two features of the table deserve attention. The first is that the two studies using a single narrow humor task, Howrigan and MacDonald with Raven's alone and Kellner and Benedek with a fluid measure, produce the smallest correlations, which is what measurement error does to an estimate. The second is that the last row is not measuring the same thing as any row above it, and that is the subject of its own section below.

Howrigan and MacDonald's regression is worth extracting because it addresses the obvious confound. They entered ten predictors at once: general intelligence, all five Big Five traits, age, height, weight, and semesters in college. Two survived. General intelligence had a standardized beta of .20 at p = .010 and extraversion had .16 at p = .037. Openness, conscientiousness, agreeableness, and neuroticism did not predict rated funniness once intelligence was in the model. The relationship between measured ability and trait structure generally is covered on the personality page.

6 Why the Association Is Verbal Rather Than Fluid

The tilt toward verbal ability is not a quirk of the samples; it follows from what a joke requires. Writing a caption means retrieving candidate meanings for the objects in the picture, finding a second frame in which those meanings mean something else, and compressing the collision into one line. Every step of that is retrieval and manipulation of stored knowledge.

Consider what the fluid tasks in these studies actually demand. Paper folding, letter sets, series completion, and matrix reasoning all present novel material with no prior knowledge required, and reward finding the rule. That is a real ability and it is the backbone of most composites, but it is not the ability that supplies the second meaning of a word. Vocabulary tests, by contrast, are direct samples of exactly the store a joke draws on. When Greengross and Miller found .38 for vocabulary against .25 for Raven's and confirmed the gap at z = 2.1, they were finding a difference their materials predicted.

Christensen and colleagues added the piece that makes the mechanism concrete. Broad retrieval ability, measured with tasks such as naming as many animals or as many things that are hot as possible in a fixed time, kept a unique effect of .20 on humor production even after general ability was modeled, while fluid reasoning kept nothing. Retrieval fluency is the ability to search semantic memory quickly and strategically, which is the operation a punchline needs most and the one a matrix problem needs least.

The distinction between reasoning with unfamiliar material and drawing on accumulated knowledge is the oldest division in the ability literature, and it is set out on the fluid versus crystallized page. The specific tasks that sample the verbal side, and how they differ from each other, are described on the verbal ability page and in the subtest descriptions for Vocabulary, Similarities, and Synonyms.

7 The Dark Humor Headline and What the Study Found

The most widely circulated claim in this area comes from one study of 156 people that measured comprehension of cartoons, not the ability to be funny, and the press coverage overstated it in three separate ways. Willinger and colleagues published Cognitive and emotional demands of black humour processing in Cognitive Processing in 2017, volume 18, pages 159 to 167.

Here is what the study did. One hundred and fifty-six adults, 76 women and 80 men, with a mean age of 33.4 and a standard deviation of 11.9, rated 12 black humour cartoons drawn from a collection by Uli Stein. Six of the cartoons dealt with death, three with physical handicap, two with disease, and one with medical treatment. Participants rated each cartoon on difficulty, fit, preference, and comprehension. Verbal ability was measured with a German vocabulary test and what the paper calls nonverbal IQ was measured with the Zahlen-Verbindungs-Test, a number connection task.

A cluster analysis split the sample into three groups. Group one, 41 people, had a verbal mean of 101.0 and a nonverbal mean of 97.8. Group two, 50 people, had 96.8 and 102.8. Group three, 65 people, the group with the highest preference for and comprehension of the cartoons, had 109.7 and 118.1. The differences were significant, F(2,153) = 112.6 for the nonverbal measure and F(2,153) = 19.9 for the verbal one, both at p at or below .0001. Group three also contained significantly more higher educated participants at p = .014.

Now the three overstatements. First, the study is about comprehension and preference, not production; nobody in it wrote a joke. Second, the group means are not high in absolute terms. A group mean of 118 on a speeded number connection task is High Average, not gifted, and it is a group average rather than a description of any individual. Third, the task labeled nonverbal intelligence is a trail making test in which the participant connects numbers in sequence as fast as possible, developed by Oswald and Roth within the mental speed tradition. It is much closer to a measure of processing speed than to a measure of reasoning, so reporting the result as a difference in nonverbal intelligence claims more construct coverage than the instrument delivers.

8 People Are Poor Judges of Their Own Jokes

Asking someone whether they are funny produces a number that barely relates to whether other people find them funny, which is why no serious study in this literature uses self-report as the outcome. The best evidence for this is Silvia, Greengross, Cotter, Christensen, and Gredlein, published in Journal of Research in Personality in 2021, volume 92, article 104089.

Across seven studies with 1,133 adults, participants created funny ideas, rated the funniness of their own responses, and had those same responses independently rated by judges. The within-person correlation between the self ratings and the judge ratings was .13, with a confidence interval from .07 to .19 across all seven studies. The authors describe this as small but significant, meaning people have some insight into which of their own attempts landed, and not much.

The paper also found that personality predicts self-rating rather than performance. Extraversion correlated .12 with rating one's own responses as funnier, interval .07 to .18. Openness correlated .09, interval .03 to .15. Women rated their own responses as less funny than men rated theirs, with a d of minus 0.28, interval minus 0.37 to minus 0.19. None of those associations is with judged funniness. They are associations with confidence.

One further result cuts against the usual story. The well known finding that most people rate themselves as funnier than average did not appear when people rated specific attempts. Asked about a particular joke they had just written, participants were relatively modest and self-critical. The inflation lives in the global self-judgment, not in the local one.

That pattern generalizes well beyond humor. Self estimates of ability in general correlate loosely with measured ability, which is the reason a self-assessment is not a substitute for an administration. The evidence on that gap is collected on the page on estimating your own ability.

9 What the Comedians Data Adds

One study went to the people who do this for a living, and the result is consistent with the student data while being much smaller and much harder to interpret. Greengross, Martin, and Miller published their comparison of professional stand-up comedians with college students in Psychology of Aesthetics, Creativity, and the Arts in 2012, volume 6, pages 74 to 82.

Thirty-one professional comedians were compared with 400 college students on the Big Five, the Humor Styles Questionnaire, a humor production task, and a measure of verbal intelligence. The comedians scored higher than the students on verbal intelligence, on humor production ability, and on each of the four humor styles the questionnaire measures. Within the comedian group, openness, agreeableness, and extraversion correlated positively with affiliative humor, and intelligence correlated negatively with self-defeating humor.

Take the sample size seriously before taking the result seriously. Thirty-one people is enough to detect a large group difference and not enough to estimate a correlation within the group with any precision. A within-comedian correlation from 31 people has a confidence interval wide enough to include almost any moderate value, so the personality findings inside that group should be read as suggestive rather than established.

There is also a selection issue that no design can remove. Professional comedians are people who tried comedy, were good enough to keep going, and stayed in the profession long enough to be recruited for a study. Any comparison between them and undergraduates confounds ability with survivorship, motivation, and years of deliberate practice. The finding that they score higher on a verbal measure is compatible with the ability being a cause, with the practice being a cause, and with both.

The honest summary is that the comedian data is consistent with the student data and adds little independent weight. What it does add is a reminder that the constructs in this literature are separable: humor production ability, humor styles, and professional success at comedy are three different variables, and the study measured all three because they do not reduce to one another.

10 The Sex Difference and the Meta-Analysis About It

The one meta-analysis in this area is about sex differences in humor production rather than about humor and intelligence, and the distinction is worth stating because the two are often confused. Greengross, Silvia, and Nusbaum published it in Journal of Research in Personality in 2020, volume 84, article 103886. A random effects model across the available studies produced a combined effect size of d = 0.321, with men's humor output rated funnier on average by judges.

An effect of about a third of a standard deviation is a real difference in means and a poor basis for any judgment about an individual. Two distributions separated by d = 0.32 overlap across roughly 87 percent of their range. Picking one man and one woman at random, the man's rated humor would be higher about 59 percent of the time, which is barely distinguishable from a coin flip. Population level effects of this size are genuine and individually invisible, which is the same arithmetic that governs every group comparison in the ability literature.

The meta-analysis has been publicly criticized on methodological and conceptual grounds, including how the underlying studies defined and scored humor production. That criticism does not make the pooled estimate disappear, and it does not settle the question either. What it means is that a reader should treat d = 0.321 as the best available pooled figure from a contested literature rather than as a fixed constant.

The primary studies show the same sex pattern in the same direction. Greengross and Miller reported d = 0.38 in their 400 person sample. Howrigan and MacDonald reported that general intelligence did not interact with participant sex in predicting rated humor, meaning the ability association held equally in both halves of their sample even where the mean differed.

11 The Mechanism Under the Joke

Humor comprehension has a documented two stage structure, and both stages draw on capacities that cognitive tests sample directly. Vrtička, Black, and Reiss reviewed the imaging evidence in Nature Reviews Neuroscience in 2013, volume 14, pages 860 to 868. Their account describes a core network in which temporo-occipito-parietal regions detect and resolve incongruity, the mismatch between what was expected and what arrived, after which mesocorticolimbic reward structures and the amygdala produce the response people experience as mirth.

Detection and resolution are the cognitive half, and they map onto measurable operations. Detecting an incongruity requires having built an expectation, which means holding the setup in working memory while the punchline arrives. Resolving it requires finding a second interpretation under which the incongruity makes sense, which means searching semantic memory for an alternative frame. That search is exactly the retrieval operation Christensen and colleagues measured with fluency tasks and found to carry a unique effect.

• Semantic retrieval. Finding the second meaning requires searching stored knowledge quickly and selectively, which is what broad retrieval ability measures and what vocabulary tests partly index.
• Incongruity detection. Noticing the mismatch requires an expectation to have been formed and held, which is a working memory operation.
• Verbal fluency and compression. Getting the collision into a single line under time pressure draws on the same machinery a verbal reasoning subtest samples.
• Selection. Most candidate jokes are bad, so producing a good one requires generating several and discarding most, which is a control operation rather than a knowledge one.

This is why the ability association shows up where it does. A test that samples verbal knowledge and working memory is sampling two of the four operations above, which is enough to produce a correlation around .4 and not enough to produce one around .8. The specific subtests that load on working memory are described on the Digit Span page and the Alphanumeric Sequencing page, and how memory relates to measured ability more broadly is on the memory page.

Nothing in the imaging account licenses a causal claim in either direction. It describes what the brain does while processing a joke, not what makes one person better at it than another. Correlation between rated funniness and vocabulary is consistent with shared verbal machinery, with shared exposure to language, and with practice at both, and the studies here cannot separate those.

12 What a Cognitive Test Can and Cannot Tell You Here

No cognitive battery measures humor, and this one does not either. What a battery can do is measure the abilities that the humor literature identifies as carrying the association. That is a narrower and more useful claim than the one the popular coverage makes.

The ability most implicated is verbal knowledge. ACIS reports it as the Verbal Comprehension Index, built from five subtests: Antonyms, Vocabulary, Information, Synonyms, and Similarities. In the published technical manual the index has an omega reliability of .9745, a standard error of measurement of 2.40 points, and a g loading of .864. The subtest g loadings within it run from .865 for Similarities and .860 for Antonyms down to .828 for Synonyms and .814 for Information. Fluid reasoning is reported separately as the FRI, with an omega of .9727 and a g loading of .922, and the two indices correlate .801 with each other, which is why separating them requires modeling rather than eyeballing.

That .801 figure is the reason the Christensen bifactor result matters so much. When two indices correlate that highly, a raw correlation between either one and an outcome is mostly carrying shared general variance. Only a model that separates the general factor from the specific ones can say whether the verbal part is doing independent work, and in the humor case it turned out that the fluid part was not. The full index structure and the six CHC domains are described on the CHC model page and on the cognitive domains page, with the reliability machinery explained on the reliability and validity page.

What a result does not tell you is whether you are funny. A high VCI is consistent with the profile the humor production studies describe and predicts nothing about any individual's ability to write a caption, because a correlation of .49 leaves roughly three quarters of the variance unexplained. ACIS is an unsupervised online assessment, not a clinical instrument, and it is not appropriate for diagnosis, hiring decisions, accommodations, or society admission.

13 Sources Behind This Page

Every number on this page comes from one of the papers below, and where a figure was taken from a later paper's report of an earlier one rather than from the original, the text says so. That distinction matters in a small literature where the same handful of correlations get passed forward across citations.

• Greengross, G., and Miller, G. (2011). Humor ability reveals intelligence, predicts mating success, and is higher in males. Intelligence, 39, 188 to 192. N = 400, r = .38 for vocabulary and .25 for Raven's, z = 2.1 for the difference, d = 0.38 for the sex difference.
• Christensen, A. P., Silvia, P. J., Nusbaum, E. C., and Beaty, R. E. (2018). Clever People: Intelligence and Humor Production Ability. Psychology of Aesthetics, Creativity, and the Arts, 12, 231 to 241. N = 270, latent correlations of .49, .38, and .22, and a bifactor model in which the fluid effect falls to .09. A copy of the accepted manuscript is available as a PDF.
• Howrigan, D. P., and MacDonald, K. B. (2008). Humor as a Mental Fitness Indicator. Evolutionary Psychology, 6, 652 to 666. N = 185, r = .29, regression beta of .20 for intelligence and .16 for extraversion. The full text is available as a PDF.
• Kellner, R., and Benedek, M. (2017). The Role of Creative Potential and Intelligence for Humor Production. Psychology of Aesthetics, Creativity, and the Arts, 11, 52 to 58. 151 participants writing punch lines for six cartoons. The correlations of .37 and .13 quoted here are as reported by Christensen and colleagues in 2018.
• Willinger, U., and colleagues (2017). Cognitive and emotional demands of black humour processing. Cognitive Processing, 18, 159 to 167, open access at PubMed Central. N = 156, three clusters, group means of 109.7 verbal and 118.1 nonverbal in the highest group.
• Silvia, P. J., Greengross, G., Cotter, K. N., Christensen, A. P., and Gredlein, J. M. (2021). If You're Funny and You Know It. Journal of Research in Personality, 92, 104089. Seven studies, n = 1,133, self to judge correlation of .13.
• Greengross, G., Silvia, P. J., and Nusbaum, E. C. (2020). Sex differences in humor production ability: A meta-analysis. Journal of Research in Personality, 84, 103886. Combined d = 0.321.
• Greengross, G., Martin, R. A., and Miller, G. (2012). Personality traits, intelligence, humor styles, and humor production ability of professional stand-up comedians compared to college students. Psychology of Aesthetics, Creativity, and the Arts, 6, 74 to 82. 31 comedians against 400 students.
• Vrtička, P., Black, J. M., and Reiss, A. L. (2013). The neural basis of humour processing. Nature Reviews Neuroscience, 14, 860 to 868. The incongruity detection and resolution account.
• Feingold, A., and Mazzella, R. (1991). Psychometric intelligence and verbal humor ability. Personality and Individual Differences, 12, 427 to 435, and Masten, A. S. (1986). Both are quoted here as reported by Howrigan and MacDonald in 2008 rather than from the originals.

Two limits belong on the record. The samples behind almost every figure above are university students in the United States, which restricts the range of both ability and age and inflates nothing but does narrow generalization. And no study cited here is longitudinal or experimental, so nothing on this page establishes that verbal ability causes funniness rather than the two sharing a common source. Correlational evidence supports a description of who tends to be funnier, not a mechanism for making anyone funnier.

The professional framework for interpreting any of this is explicit. The International Test Commission guidelines on test use and the Standards for Educational and Psychological Testing (2014), published jointly by the American Educational Research Association, the American Psychological Association, and the National Council on Measurement in Education, require that score interpretation be tied to documented evidence for the specific use, that classification be reported alongside its measurement error, and that the limits of the reference sample be disclosed. APA standards on test use impose the same discipline on anyone reporting a result: state what the evidence supports, state what it does not, and never let a correlation about a group become a statement about a person.

14 Frequently Asked Questions