Can candidates cheat psychometric tests with AI — and what should employers do about it?

Generative AI has changed unproctored online testing. Some test types are far more exposed than others, and the right response is a combination of test design, detection and policy — not abandoning online assessment.

Which test types are exposed?

Not all assessments are equally vulnerable. Verbal reasoning items — a passage and a set of statements — are exactly the kind of task large language models handle well. Knowledge questions are trivially answerable. At the other end of the scale, timed tests with short item-level time budgets leave little room to consult anything, and game-based assessments measure response processes that are hard to delegate to another tool at all.

Personality questionnaires sit in a different category: there is no "correct" answer to look up, but candidates can ask an AI what profile a role probably rewards. That risk predates AI — coached responding has always existed — and the countermeasures are the same: social desirability scales, response-consistency checks and norm-referenced interpretation.

What detection looks like

Detection works on converging evidence rather than a single signal. Response times tell you when answers arrive faster than the item can be read. Answer patterns — accuracy that is flat across difficulty levels, for example — look nothing like unaided human performance, where accuracy falls as items get harder. Copy protection on test pages raises the cost of pasting items into another window. None of these alone is proof; together they identify sessions that merit a second look.

The decision that follows matters as much as the detection. Flagged sessions should be reviewed by a human, with the evidence visible, and the candidate given a route to respond — a supervised retest is often the fairest outcome. Automatic rejection on an algorithmic flag is both poor practice and, in many jurisdictions, a legal risk.

Designing tests that are harder to cheat

The strongest protection is built into the test itself. Item banks with randomised selection mean no two candidates sit the same form, so leaked items lose value quickly. Short per-block time limits compress the window for outside help. And verification testing — a short supervised retest for shortlisted candidates — makes the unproctored stage self-correcting: a large gap between the two scores is itself informative.

Continuous validation closes the loop. Monitoring norm drift tells you whether scores are creeping upward in ways that ability distributions can't explain — the population-level signature of assisted responding.

What employers should do now

Be explicit with candidates. State in the invitation that AI tools are not permitted, that activity is monitored, and that results may be verified. Most candidates are honest; clear rules keep them that way, and clear notice is what makes enforcement defensible.

Fairness has to survive the response. Approved assistive technology and agreed reasonable adjustments must keep working — a screen reader is not a cheating tool — and any detection method should be validated to show it does not disadvantage disabled candidates or any other group.