Behavior Hiring
Evidence review

Can candidates game hiring in the age of AI?

An evidence review of applicant faking, application volume and generative AI in hiring: what the faking literature establishes, what AI changed, and which methods still hold up under stakes.

Short answer

Applications per job opening rose from 116 in 2022 to 244 in 2025, a 111% increase, while the number of recruiters per organization fell 56% over the same period. Faking on self-report measures is well documented and predates AI. Estimates of how many applicants distort responses have risen from around 14% in the 1960s to 30%, and more recently to about 62%.

Key findings
  • Applications per job opening rose from 116 in 2022 to 244 in 2025, a 111% increase, while the number of recruiters per organization fell 56% over the same period.
  • Faking on self-report measures is well documented and predates AI. Estimates of how many applicants distort responses have risen from around 14% in the 1960s to 30%, and more recently to about 62%.
  • Personality measures predict performance far more weakly among hired applicants (mean r = .08) than among incumbents (.28), which is what you would expect if high-stakes self-reports are distorted.
  • Faking does not merely shift everyone up. It changes who gets hired: applicants who scored high at hiring have scored materially lower on retest months later.
  • Across matched samples, criterion validity is about a third lower when the same assessment is taken in a high-stakes setting than a low-stakes one — and the loss concentrates in the top of the score range, the only part employers hire from.
  • Faking ability is itself predicted by cognitive ability, so distortion is not noise spread evenly across a candidate pool. It is an advantage held by some candidates and not others.
  • Generative models can produce near-ideal personality profiles on high-stakes assessments, including on forced-choice formats designed to resist faking.

Yes, and they always could. What has changed is the cost. Writing a polished, keyword-matched application used to take an evening; now it takes a minute, and the same tools will rehearse a candidate through the interview that follows. The useful question is not whether applicants use AI. It is which parts of a hiring process were resting on effort as a filter, and what is left once that filter is gone.

What the volume data shows

Greenhouse's benchmark report, drawing on more than 640 million applications across over 6,000 companies, gives the clearest picture of the squeeze.

2022 2025 Change
Applications per job 116 244 +111%
Applications per recruiter 1,610 4,890 +204%
Recruiters per organization 10.43 4.62 −56%

Twice the applications, half the recruiters. Whatever screening judgment was being applied in 2022, less of it per application is being applied now.

It is worth being careful about causation. Application volume rose for several reasons at once, including a softer labour market and one-click applications, and this data does not isolate how much of the increase is AI-assisted. What it does establish is the condition: more paper arriving at a thinner filter.

Why the résumé was already weak

The common reaction is that AI has broken résumé screening. The more accurate reading is that résumé screening was never strong and is now visibly weak.

Years of job experience, the single most common thing a résumé is screened on, has a validity of .07 against job performance. That was true before anyone could generate an application in seconds. What AI changed is that the weak signal is now also uniform: when every application is well-written and keyword-matched, the differences a screener used to read, effort, care, specificity, stop varying.

That has a second-order effect worth naming. If polish no longer distinguishes candidates, screeners fall back on whatever still varies, which tends to be pedigree: the school, the recognizable employer, the familiar job title. Those are weak predictors and carry adverse impact. A process that loses its ability to read effort does not become neutral. It becomes more dependent on proxies.

What the faking research establishes

The interview-coaching question is really the old faking question in a new costume, and that literature is substantial.

Applicants score higher than incumbents on positively valued traits, consistently, across meta-analyses. Estimates of prevalence have climbed over sixty years, with more recent organizational estimates around 62%, though these figures vary with how faking is defined and measured, and should be read as a range rather than a count.

The finding with the sharpest practical edge is from Jeong and colleagues: personality-performance correlations average .08 among hired applicants versus .28 among incumbents. The same instrument predicts substantially less well in exactly the setting it is sold for.

And faking is not harmless inflation. Griffith and colleagues retested people on a conscientiousness measure six months after hire and found some who had scored high at application scored considerably lower afterward. Using the later scores would have produced different hiring decisions. The distortion does not lift all boats; it reorders them, and the people it reorders downward are not randomly distributed, they are the ones who answered honestly.

The 2025 critical review that settles most of this is Dunlop, Xia and Anglim in Current Opinion in Psychology (doi.org/10.1016/j.copsyc.2025.102057). Two findings in it matter more than the prevalence estimates. First, a meta-analysis of matched samples who took the same assessment under low- and high-stakes conditions found criterion validity in the high-stakes condition was about a third lower, across a range of criteria. Faking does not just add noise; it costs you a third of what the instrument was worth.

Second, and worse for anyone hiring: validity is not evenly damaged across the score range. Using segmented regression, Christiansen and colleagues found that for real job applicants, the criterion validity of personality scores is significantly lower in the upper part of the score range than in the lower part — a pattern absent in a low-stakes incumbent sample. The top of the list is where fakers cluster, and the top of the list is the only part you hire from. The instrument works best exactly where you are not looking.

Faking is also not randomly distributed, which is the part most discussions miss. Grieder and colleagues separated how much people fake from how well they can fake, and found that faking ability is substantially predicted by cognitive ability — crystallised intelligence in particular (R² = .39 for the pattern of faking, against .02 for sheer elevation). So distortion is not a fog over the whole candidate pool. It is an advantage, and it accrues to a group that overlaps with the group the test was supposed to identify. Birkeland and colleagues add the final turn of the screw: the two traits most predictive of job performance, conscientiousness and emotional stability, are the two faked hardest.

On AI specifically, the evidence is no longer only inferential. Phillips and Robie tested whether generative models could produce ideal personality profiles on high-stakes assessments and found they could, including on forced-choice formats built specifically to resist faking. Lievens and Dunlop describe the emerging case as applicants using AI as a substitute test-taker rather than as a coach. How well the model does it depends partly on how it is prompted, which means the advantage tracks prompting skill — another ability unrelated to the job.

What holds up

The pattern across all of this is consistent. Methods that ask a candidate to describe themselves degrade under stakes. Methods that ask for a specific, verifiable account of something they did, and then probe it, degrade much less.

A generated résumé can claim ownership of a project. It cannot, under three follow-up questions, produce who was in the room, what the disagreement was about, what the candidate personally said, and how it turned out, with the internal consistency that a real memory has. Coaching helps a candidate deliver a strong first answer. The evidence tends to live two or three probes down, which is also where coaching runs out.

This is the practical argument for structure, and it is now stronger than it was. Structured interviews already had the highest validity of any method studied. In an environment where the paper signal is both weaker and more uniform, the gap between what a résumé tells you and what a well-run interview tells you is widening.

What to do with it

Stop screening on things AI writes well. Keyword matching and polish no longer discriminate. Screen on the few genuine non-negotiables and move people to a structured conversation sooner.

Push past the first answer. Budget fewer questions with more probing rather than more questions. A rehearsed answer and a real one are hard to tell apart at the surface and easy to tell apart underneath.

Ask for specifics only a participant would have. Not "tell me about a project you led" but what the constraint was, who objected, what you did when the first approach failed.

Do not treat AI use as misconduct in itself. A candidate using AI to write a clear application is doing what the tools are for, and detection tools are unreliable enough that acting on them creates its own risk. The sound response is a process that is robust to it, not a policy against it.

Method

A narrative review. The volume figures are from Greenhouse's published benchmarks and describe their customer base rather than the whole labour market. The faking findings are from the peer-reviewed literature on self-report measures in selection, which long predates generative AI; the validity-decline, score-range and faking-ability findings are from primary sources cited in the text, chiefly the 2025 critical review in Current Opinion in Psychology. The claim that generative models can fake these assessments is directly measured (Phillips & Robie, 2024); what remains reasoned rather than measured is how widely applicants are currently doing it, and what that does to validity in live hiring. Claims about current AI use by candidates circulate widely in vendor surveys with variable methodology, and this review does not rely on them.

Sources

  1. Dunlop, P. D., Xia, L., & Anglim, J. (2025). Applicant faking and high-stakes assessment: a critical review. Current Opinion in Psychology. https://doi.org/10.1016/j.copsyc.2025.102057
  2. Greenhouse. Hiring benchmark report: applications per job, applications per recruiter and recruiters per organization, 2022–2025.
  3. Jeong, S., et al. Personality–performance relationships among hired applicants versus incumbents (mean r = .08 versus .28).
  4. Griffith, R. L., et al. Retest of applicant conscientiousness scores six months after hire.
  5. Christiansen, N. D., et al. Segmented regression of personality score range and criterion validity in real job applicants.
  6. Grieder, S., et al. Separating faking amount from faking ability, and the prediction of faking ability by crystallised intelligence (R² = .39 versus .02).
  7. Birkeland, S. A., et al. Meta-analysis of applicant faking by trait: conscientiousness and emotional stability faked hardest.
  8. Phillips, J., & Robie, C. (2024). Generative AI and ideal personality profiles on high-stakes and forced-choice assessments.
  9. Lievens, F., & Dunlop, P. D. Applicants using AI as a substitute test-taker rather than as a coach.
  10. Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology, 107(11), 2040–2068. https://doi.org/10.1037/apl0000994

How to cite this page

Behavior Hiring. (2026). Can candidates game hiring in the age of AI?. https://behaviorhiring.com/research/candidate-faking-ai-hiring/

Monthly Research Brief

One email a month. New research, a practical tool, and what changed in the field.

No spam. Unsubscribe any time.