The Impact of Impression Management on Writing Tendencies: Investigating Faking Susceptibility of Open-Ended Personality Assessments

• 成果类型: Article; Early Access
• 署名作者: Speer, Andrew B.; Chawota, Takudzwa A.; Rudolph, Cort; Qaqish, Amani
• 署名单位: Indiana University System; Indiana University Bloomington; Missouri State University; Wayne State University
• 刊物名称: JOURNAL OF APPLIED PSYCHOLOGY
• ISSN/ISSBN: 0021-9010
• DOI: 10.1037/apl0001397
• 发表日期: 2026
• 关键词: situational judgment tests; PERSONNEL-SELECTION; applicant faking; job interviews; CLASSIFICATION; deception; language; BEHAVIOR; domains; traits

摘要

Applicant faking on self-report personality measures has long been a concern. One method that might be more resistant to faking is open-ended personality assessments. Guided by a new framework of open-ended faking behavior, we investigated how faking emerges in open-ended personality assessments. Additionally, we examined an alternative form of faking: artificial intelligence (AI)-generated responses. Results showed that elaborate responding (i.e., effortful crafting of narratives, Cohen's d = .29) and selective information management (i.e., promoting positive information while withholding negative information, d = .72) were higher in faked than honest conditions and that individuals with higher trait levels were more capable of producing elaborate responses and providing favorable information under motivated conditions. When faking, scores increased more on Likert self-reports (d(-) = .65) than on open-ended personality scores ( d & strns; = .26) derived via natural language processing, though when trained to predict human ratings rather than honest Likert scores natural language processing scores were more susceptible to faking ( d & strns; = .63). Training sample configuration also influenced psychometric properties, such that training on a mixed sample of honest and motivated responses yielded the best results. Finally, AI-generated responses (fully synthetic and human-AI hybrid) received scores that varied considerably across models, though they typically were higher than human scores. However, a customized algorithm accurately differentiated human from AI responses with near-perfect accuracy (across seven large language models and for humans using AI), with only a 1% false positive rate. Collectively, these findings suggest that open-ended personality assessments, when paired with natural language processing scoring, have potential to offer a scalable, more faking-resistant approach to personality assessment.