Intelligence tests are a weaker predictor of job performance than long believed, large study finds

A large synthesis of recent workplace data reinforces evidence that general intelligence tests are a weaker predictor of overall job performance than long believed. By combining data from over 109,000 workers across more than 200 contemporary samples, researchers found that the predictive power of cognitive ability tests is less than half of the historical benchmark widely cited in textbooks. The findings, published in the Journal of Business and Psychology, suggest that employers should rely on a broader mix of hiring assessments rather than treating mental ability tests as a standalone measure of future success.

General mental ability refers to a person’s broad capacity to learn, reason, process information, and solve unfamiliar problems. Rather than representing a single narrow skill, it reflects the shared underlying capability that links specific talents like verbal comprehension, math skills, and spatial reasoning. In the workplace, standardized tests measuring this general ability are often used during hiring to estimate how well an applicant will learn job duties and perform once hired.

The traditional view in workplace psychology was that general mental ability stood out as the single best predictor of job performance across all types of occupations. This view was cemented by an influential 1998 review that synthesized decades of hiring data and estimated the correlation between mental ability and job performance to be just over .50. For a positive relationship like this one, a correlation of 0 indicates no relationship and 1 a perfect one, making .50 a strong link by the standards of behavioral research.

In recent years, that benchmark has come under scrutiny due to how earlier researchers handled a statistical issue known as range restriction. When a company hires only top-scoring applicants, the pool of hired workers has a much narrower range of test scores than the general pool of job seekers. Because looking only at hired workers can artificially shrink the apparent link between test scores and job performance, researchers apply mathematical adjustments to estimate how the test would perform across all applicants.

The authors of a 2022 meta-analysis, led by Paul Sackett, a co-author of the new study, argued that earlier workplace studies had applied overly aggressive statistical adjustments for range restriction, which inflated the apparent value of cognitive tests. When they recalculated those older datasets using more conservative adjustments, the estimated correlation dropped from .51 to .31, a conclusion other researchers have disputed. The foundational databases also relied mostly on studies conducted before the mid-1980s, many of them involving manufacturing jobs.

To see how these tests perform in modern workplaces, researchers recently conducted three separate meta-analyses looking at studies from 1990 through 2022, including a 2024 longitudinal meta-analysis that documented a sharp drop in the predictive power of cognitive tests over recent decades. Although all three modern reviews yielded lower correlations ranging from .16 to .23, they drew from partially overlapping time periods and shared several of the same underlying studies. The new research, led by Saron Demeke, a PhD candidate in industrial-organizational psychology at the University of Minnesota – Twin Cities, sought to merge these three datasets, remove all duplicate studies, and establish an accurate modern estimate of how well general mental ability predicts job performance.

“For decades, general mental ability (GMA), or cognitive ability, was described as the single best predictor of job performance, but that conclusion rested largely on studies conducted before 1980, and the nature of work has changed substantially since then,” Demeke told PsyPost. “Three recent meta-analyses, or comprehensive reviews of existing studies, reported much lower estimates. These papers, including work by members of our team, drew on partially overlapping sets of studies, so their results could not simply be averaged.”

Demeke and colleagues also wanted to re-evaluate the long-held assumption that intelligence tests become better predictors of performance as a job becomes more mentally demanding. The team gathered the full datasets from the three recent meta-analyses, which covered studies conducted between 1990 and 2022, and systematically cross-referenced the databases to strip out overlapping samples. To make the datasets comparable, they also dropped studies that used grade point averages instead of a test, relied on self-, peer-, or subordinate ratings, measured performance during training, or used data collected before 1990.

This filtering process left a combined dataset of 212 independent samples encompassing 109,396 workers. Job performance in these studies was primarily measured through supervisory ratings of overall performance or specific task proficiency, alongside a smaller number of objective performance metrics. To ensure their conclusions did not depend on one debated statistical choice, the researchers analyzed the data using multiple types of range restriction adjustments and different methods for weighting the three datasets.

Before applying any statistical adjustments for range restriction or measurement error, the average observed correlation between general mental ability and job performance across the 212 modern samples was .14, compared to .24 in older databases built mostly from studies conducted before the mid-1980s. After applying statistical corrections to account for restricted applicant pools and imperfect supervisor ratings, the contemporary correlation fell between .18 and .22 across all tested methods. The authors proposed .20 as the single best summary estimate for modern workplaces, a figure less than half the size of the historical .51 benchmark and noticeably lower than the .31 revised estimate for older data.

The authors favored the equally weighted figure of .20 partly because the largest of the three datasets, covering about 65,000 workers, came from studies of a single commercial test, the Wonderlic, and they did not want one test to dominate the estimate. They described the choice as a judgment call.

“As an illustration, if cognitive ability scores were used to make hiring decisions, someone scoring in the top 20% on a cognitive ability test would be roughly 1.6 times as likely to be rated as an above-average performer as someone scoring in the bottom 20%,” Demeke explained.

Demeke said two results stood out to the team. “First, the estimates barely moved (between .18 and .22) after we removed the overlapping studies and tried different statistical adjustments, which correct for things like measurement error and the limited range of scores among people who get hired.”

To investigate how job complexity affects these numbers, the researchers examined a subset of 137 single-occupation samples that could be matched to the Occupational Information Network, a comprehensive federal database of job characteristics. Each occupation was scored on three dimensions of complexity: cognitive demands involving information processing (“Data”), interpersonal demands involving teamwork and communication (“People”), and manual or technical demands involving equipment and physical processes (“Things”). The team then ran statistical models to see if these job ratings predicted how well cognitive tests worked in different roles.

Contrary to traditional expectations, cognitive complexity alone showed virtually no relationship with how well mental ability tests predicted job performance, yielding a correlation of -.04. Interpersonal complexity also showed no meaningful link to test validity. In contrast, manual and equipment-related complexity showed a positive relationship with test predictive power, with a correlation of .29.

When the researchers looked at how these job demands interacted with one another, they found that cognitive tests tended to be most predictive in jobs that combined high levels of both cognitive (“Data”) and manual (“Things”) demands. In the occupational database, high “Things” scores go beyond routine physical labor to include tasks like inspecting equipment, troubleshooting mechanical failures, and controlling complex machinery. The findings suggest that general mental ability remains particularly relevant in dynamic environments where workers must process complex information to solve hands-on, technical problems.

This pattern matched an earlier analysis by members of the research team but differed from the 2024 longitudinal meta-analysis, which had linked hands-on job demands to weaker predictive power. The authors suggest that the effect of hands-on demands may depend on whether a job also requires heavy information processing.

“Second, we expected the link to be strongest in the most mentally demanding jobs, but what we found was more nuanced,” Demeke said. “The relationship tended to be strongest in jobs that pair thinking with hands-on work, like troubleshooting, inspecting, and operating equipment. One possible explanation is that ‘hands-on’ work in many jobs today involves more than just routine physical tasks and often requires substantial information processing and problem solving.”

Emily A. Willoughby, a research assistant professor in the Department of Psychology at the University of Minnesota and board member of the International Society for Intelligence Research, is in the same department as three of the study’s authors but was not involved in the research. She described the new paper as one that “mostly consolidates and confirms, rather than overturns” earlier work.

“The fact that intelligence test scores can predict job performance is very well-established, but the oft-repeated correlation of about .50 (which is pretty high, as psychological correlations go) is based almost entirely on studies done before 1980,” Willoughby told PsyPost. “There is reason to suspect that those old correlations were too generous, and Demeke et al. wanted to look only at newer data. The workforce today is clearly different than it was prior to 1980, and it’s reasonable to assume that older data doesn’t capture the modern association between cognitive ability and job performance. And when looking only at studies conducted after 1990 (and count each study once), the authors find the relationship to be closer to about .20, an estimate that seems to be remarkably resistant to correction method, weighting, and so on.”

To account for why cognitive ability tests appear less predictive of overall job performance today than half a century ago, the researchers point to two likely explanations, though the study did not directly test them. First, the nature of work has shifted away from individual manufacturing roles toward service and knowledge jobs that heavily emphasize teamwork, communication, and adaptability. If supervisors now factor these social and emotional skills into overall ratings, cognitive ability would represent a smaller piece of what they are judging.

Second, the pool of people applying for any given job appears to have become more similar in cognitive ability. The 2024 longitudinal meta-analysis found that the spread of test scores within job-specific applicant pools shrank by about 22 percent between data collected before 1973 and after 2004, which its authors attributed partly to better career information and growing degree requirements. When applicants already have similar levels of education and ability, test scores vary less, which limits a test’s ability to separate top performers from average ones. The authors stress that this is a real loss of predictive power for employers, not a statistical distortion to be corrected away.

“The most important takeaway is that general mental ability may not relate as strongly to job performance as the field has long assumed to be true,” said Erik Gonzalez-Mulé, a professor of organizational behavior and human resources, the Randall L. Tobias Chair in Leadership, and chair of the Kelley Direct Programs at Indiana University’s Kelley School of Business, who was not involved in the study. “The authors attribute that to a) the changing nature of work since the foundational studies on the validity of GMA were conducted, and b) overly generous range restriction corrections.”

He sees the findings as a challenge to what the field has long taught. “It definitely challenges the long-standing view in the literature that GMA is among the, or the, most important individual differences variable that predicts job performance,” Gonzalez-Mulé told PsyPost. “This is a big deal because one of the first things we would teach in our courses or tell practitioners about selection is to use GMA as a primary tool to have the highest likelihood of selecting good employees. Now, that guidance is tempered a bit (although the validities for GMA that the paper reports are still consequential in a personnel selection setting, just not nearly as strong as previously assumed).”

As with all research, there are some caveats to consider when interpreting these results. The finding that general mental ability has a modest .20 correlation with overall job performance does not mean cognitive ability is irrelevant in the workplace. As the authors note, prior research indicates that intelligence tests remain strong predictors of how well workers perform in formal job training, with correlations ranging from .35 to .49, as well as how well they master hands-on technical proficiency.

“As we discuss in the paper, it’s important to consider the outcome that is being predicted,” Demeke noted. “Our results are about broad on-the-job performance. Our estimate is also an average across many settings, and the relationship varies between them. In most settings, we would expect the true relationship to fall roughly between .07 and .29, so it can be quite weak in some and moderate in others.”

Willoughby emphasized that general readers should not mistake a correlation of .20 for proof that intelligence no longer matters at work. “Readers might be tempted to conclude from the .50 -> .20 decline that intelligence matters little for job performance, or that intelligence stops mattering after a certain threshold,” Willoughby said. “This would be wrong: a pooled correlation of .20 is still a meaningful predictor of job performance, and a highly replicable correlation of that magnitude translates into significant material gains for a company with many employees.”

Most of the studies measured performance through supervisors’ ratings rather than objective records. In the authors’ reanalysis of one of the three datasets, studies that used objective performance measures showed a higher average correlation, about .35, while studies of managers showed the lowest, about .17.

Willoughby expressed fairly high confidence in the headline estimate but raised two cautions. “First, the criterion is mostly supervisor ratings, which are a notoriously noisy measure and are well known to be less reliable than more objective measures, such as performance, which correlates more strongly with IQ,” Willoughby said. “As such, some of the decline from .50 to .20 may reflect a change in how ‘performance’ is defined and doesn’t necessarily capture how much ability really matters.”

In addition, the job complexity analyses relied mostly on office, administrative, management, and sales occupations, with fewer samples representing industrial production, maintenance, or repair roles. Future research will need to examine a wider variety of hands-on occupations to see how well these complexity patterns hold up across the broader economy.

By convention, researchers usually require a p-value below .05 before treating a result as reliable. “Second, the key interaction (that validity increases with cognitive complexity only in jobs that require hands-on or technical work) is only marginally significant (p = .059) after accounting for occupations appearing more than once in the data,” Willoughby explained. “More research will be needed to replicate this interaction.”

Gonzalez-Mulé noted that the complexity findings are the part of the study most open to misinterpretation, and offered an example of how the two dimensions work together. “For example, a plumber would score high on ‘Things’ but relatively low on ‘Data’, whereas a trainer of plumbers who has to also educate new plumbers and keep up-to-date with best practices to effectively train new plumbers would also score high on ‘Things’ but higher on ‘Data’; the results would suggest that GMA would predict performance of the trainer to a greater degree than it would for the individual plumber,” he explained.

“As noted earlier, the amount of interaction with people (‘People’) has no effect on the validity of GMA which, combined with the increase in focus of 21st century jobs of interacting with people (either in service or teamwork contexts, etc), provides a theoretical explanation for the reduction in validity of GMA over time.”

Future research could also examine how the rapid adoption of generative artificial intelligence alters workplace demands. As artificial intelligence takes over routine information processing, the relationship between human cognitive test scores and job performance may continue to shift.

“We encourage future research to examine why the relationship is weaker today,” Demeke added. “For example, one possibility is that ‘good performance’ now includes more teamwork and interpersonal skills, which cognitive ability predicts less well; examining how overall job performance changed over time could address this.”

Gonzalez-Mulé suggested that future studies should also examine other job characteristics, such as worker autonomy, and test whether alternative hiring assessments like interviews and personality tests have experienced similar declines over recent decades. “It could be that GMA has stronger validity in jobs with more autonomy, as higher GMA people might be better able to figure out priorities and determine where to spend their efforts,” Gonzalez-Mulé said.

“Many of those selection predictor validities are also based on older data (similar to the criticism the current paper levies against GMA), so it might be useful to see how the changing nature of work has changed those validities as well and compare them to GMA. Maybe all of these relationships have gotten weaker (as the foundational theories and measures are several decades old), and there are other things that better predict performance at work that we aren’t even studying or measuring yet!”

Because the predictive power of cognitive tests differs so noticeably between older and newer studies, the authors recommend that researchers report modern findings separately rather than pooling them with older data.

Willoughby also called for directly testing the competing explanations for the decline. “Whatever supervisor ratings are actually capturing could and should be empirically measured (or use objective output measures, where they exist), to see whether the decline shows up there as well,” she said, adding that researchers should also test whether occupations with the greatest narrowing in applicant test scores show the steepest drop in predictive power. “Both of these explanations are possible, and they are not mutually exclusive, but both have very different implications: either ability really does matter less for what employers are now looking for, or ability matters as much as ever but the sorting work now occurs before hiring, e.g. in educational attainment and job requirements.”

For employers and hiring managers, the findings indicate that general mental ability tests still provide useful information, even if they are not the standalone predictors once claimed. Relying less heavily on cognitive tests can also help organizations reduce adverse impact, a term for when a hiring test disproportionately screens out applicants from certain demographic groups. Combining cognitive assessments with other job-relevant tools, such as structured interviews or work sample tests, may help organizations build selection systems that capture both technical problem-solving and interpersonal skills and are seen as fairer.

“Our findings do not suggest that GMA is unrelated to job performance, or that being able to learn quickly and solve problems isn’t useful at work,” Demeke said. “For context, it is still a better predictor of job performance than the typical informal job interview.”

“Job performance is a product of many things, including the knowledge and skills someone built up, their personality and work habits, and how well they work with others,” Demeke concluded. “Cognitive ability tells us something about how someone is likely to perform, but it also leaves much of what drives performance unexplained. GMA could be best used as one part of a broader selection system alongside things like formal or structured interviews and work samples.”

The study, “Is general mental ability still the best predictor of job performance? Integrating contemporary meta-analytic evidence,” was authored by Saron Demeke, Paul R. Sackett, Christopher D. Nye, Yujia Li, and Nathan R. Kuncel.