AI Easily Fakes Personality Tests Designed to Outsmart Humans

Generative artificial intelligence has mastered yet another human domain: acing workplace personality assessments. In a newly published study in the International Journal of Selection and Assessment, researchers found that advanced generative AI models can manipulate responses on pre-employment personality tests far more effectively than human job seekers—even on sophisticated assessment designs specifically engineered to prevent cheating.

AI Beats Humans at Gaming Job Assessments

Pre-employment personality tests have long faced a persistent challenge: applicants distorting their answers to look like ideal hires. To combat this issue, industrial psychologists developed complex assessment designs intended to make it difficult for candidates to guess the “correct” answers or inflate their scores.

However, the new study indicates that these psychometric defenses may no longer hold up against artificial intelligence. When instructed to simulate job applicants, frontier AI models demonstrated a striking ability to strategically game personality assessments, outperforming human test-takers across traditional and cheat-resistant formats alike.

Putting Anti-Faking Formats to the Test

To evaluate whether modern assessments can withstand AI manipulation, researchers evaluated 219 accounting and finance professionals alongside three frontier AI models. Human participants completed the HEXACO personality inventory under both an honest baseline condition and a simulated high-stakes job applicant scenario. In the applicant condition, candidates targeted a role requiring elevated levels of Honesty-Humility and Conscientiousness.

The investigation tested standard single-statement assessments against two types of forced-choice formats: one scored using Classical Test Theory (FC-CTT) and another using Item Response Theory (FC-IRT). Forced-choice designs require respondents to choose among statements of seemingly equal desirability, making it harder to discern the preferred answer. Meanwhile, the three AI models were repeatedly prompted with analogous instructions to simulate AI-assisted applicants competing for the same role.

Forced-Choice Formats Fail to Stop AI

The study found that two of the three AI models consistently outscored human test-takers across all assessment formats in applicant settings. The algorithms easily navigated the complex trade-offs presented by forced-choice questions to produce highly optimized, competitive candidate profiles.

Forced-choice formats also yielded mixed results among human participants. While the FC-CTT format showed less score inflation than FC-IRT for Honesty-Humility and maintained better rank-order stability under applicant conditions, this resistance did not extend to Conscientiousness. Crucially, according to the researchers, neither forced-choice format demonstrated clear improvements over traditional single-statement tests in terms of convergent or criterion-related validity, challenging the long-held assumption that forced-choice designs fundamentally reduce faking.

Implications for the Future of Hiring

As job seekers increasingly turn to generative AI tools as personal interview and test-taking coaches, traditional psychometric defenses are proving vulnerable. The authors caution that organizations cannot simply rely on forced-choice formats to prevent candidate misrepresentation in high-stakes hiring.

With AI capable of out-gaming both standard tests and human applicants, the findings underscore the urgent need for employers and organizational psychologists to rethink automated screening methods and explore new approaches to evaluating talent in the age of generative AI.

---