Personnel selection · Research companion

Three in ten applicants fake the personality test. It’s not the ones you’d guess.

Published within-person estimates put faking at roughly 30% of genuine applicants — up to one in two — and the applicants inflating their scores the most are the lower-ability ones. Under top-down selection they gain seats at the organization’s expense. This page lets you price that. The model behind it is a Monte Carlo simulation whose parameters are solved to reproduce published meta-analytic applicant evidence — not assumed.

1 · Who fakes: the compensatory pattern

Scatter plot of simulated applicants: inflation of personality scores on the vertical axis against real cognitive ability on the horizontal axis. Honest responders sit at zero inflation. Fakers form a downward-sloping cloud: lower-ability applicants inflate the most, higher-ability applicants inflate far less. — **Each dot is an applicant.** Honest responders (gray) add nothing to their scores. Fakers (red) inflate — and the lower their real cognitive ability, the more they add. Simulated pool from the manuscript’s mechanism model, calibrated so it reproduces the published applicant evidence; the pattern matches the field finding that ability correlates −.19 with measured score inflation among genuine applicants.

The fingerprint this pattern leaves in the aggregate data: the Conscientiousness × cognitive-ability correlation reverses sign when the test decides who gets hired, from r̅ = +.03 among employees tested for research to −.07 among genuine applicants, across meta-analytic databases of 130,000–188,000 test-takers. The other Big Five relations do not shift. This compensatory pattern is the only faking pattern that reproduces the meta-analytic evidence — see the full figures.

2 · Run it on your own funnel

The model below is fit to the published evidence (applicant–incumbent mean shift d = .45, Birkeland et al., 2006; the applicant correlations above; operational validities from Sackett et al., 2022). You choose the context, how many applicants fake, how selective the hire is, and what the job pays. Every output is read from precomputed simulation runs of 400,000-applicant pools — nothing is approximated in the browser.

Applicant pool context Pooled applicant evidence (r = −.07) US applicant pools (r = +.01) Non-US applicant pools (r = −.27)

Applicants who fake 30%

The Griffith & Converse (2011) synthesis concludes 30% ± 10%; within-person estimates run 30–50% (Griffith et al., 2007; Donovan et al., 2014). The manuscript's main simulations use 40%.

Selection ratio (share of applicants hired) 10%

Hires per year

Average salary of the role

of top-down hiring decisions differ from what honest responding would have produced

mean hire performance lost vs. honest responding

performance gap between each faked-in hire and the honest applicant they displaced

The adverse-impact audit sees nothing.

Dollar values are illustrative, using the convention that the standard deviation of job performance in dollars equals 40% of salary (Schmidt & Hunter, 1983). Simulation outputs, not forecasts for any specific organization.

3 · What to do about it

Stop porting research-setting numbers into hiring decisions. Predictor intercorrelations are setting-specific. A validity model built on incumbent or research-sample correlations does not describe an applicant pool under real stakes.
Use countermeasures even when your psychometrics look clean. In these simulations, reliabilities, means, and selection-rate statistics stay nearly undisturbed while 1 in 5 or more hiring decisions change. Verification warnings and forced-choice formats are warranted on prevention grounds, not detection grounds.
Measure integrity directly. Inferring it from the personality composite is exactly the channel faking corrupts. In every specification that fit the published targets, applicants who gained seats through faking were lower in integrity than those they displaced — by 0.2–0.8 SD depending on the integrity measure used as anchor.
Do not let a passed bias audit stand in for assessment health. Selection-rate impact ratios (the statistic behind NYC Local Law 144 audits) are essentially undisturbed by faking in these simulations, while the composition of who gets hired degrades.

For concrete steps — what to do this quarter and how to structure a selection system around this evidence — see the playbook.

Model inputs and assumptions

Operational validities: cognitive ability .31, Conscientiousness .19 (Sackett, Zhang, Berry, & Lievens, 2022, Table 4). Equal-weight standardized composite, top-down selection.
Applicant–incumbent Conscientiousness mean shift: d = .45 (Birkeland et al., 2006). Faking prevalence range: Griffith, Chmielowski, & Yoshita (2007); Donovan, Dwight, & Schneider (2014); Griffith & Converse (2011).
For each setting of the controls, the faking parameters (mean inflation and its ability slope) are jointly re-solved so the simulated pool reproduces the mean shift and the context's applicant correlation. Consequences are measured against the no-faking counterfactual on the same simulated pool.
Subgroup inputs for the adverse-impact ratio: d = .79 (ability; Roth et al., 2011, via Sackett et al., 2022) and −.07 (Conscientiousness; Dahlke & Sackett, 2017), minority pool share 1/6, common cutoff.
Idiosyncratic faking noise σ = .50 (the manuscript reports sensitivity runs at .25 and 1.00).