Lying on the Dissection Table: Anatomizing Faked Responses

Research has shown that even experts cannot detect faking above chance, but recent studies have suggested that machine learning may help in this endeavor. However, faking differs between faking conditions, previous efforts have not taken these differences into account, and faking indices have yet to be integrated into such approaches. We reanalyzed seven data sets (N = 1,039) with various faking conditions (high and low scores, different constructs, naïve and informed faking, faking with and without practice, different measures [self-reports vs. implicit association tests; IATs]). We investigated the extent to which and how machine learning classifiers could detect faking under these conditions and compared different input data (response patterns, scores, faking indices) and different classifiers (logistic regression, random forest, XGBoost). We also explored the features that classifiers used for detection. Our results show that machine learning has the potential to detect faking, but detection success varies between conditions from chance levels to 100%. There were differences in detection (e.g., detecting low-score faking was better than detecting high-score faking). For self-reports, response patterns and scores were comparable with regard to faking detection, whereas for IATs, faking indices and response patterns were superior to scores. Logistic regression and random forest worked about equally well and outperformed XGBoost. In most cases, classifiers used more than one feature (faking occurred over different pathways), and the features varied in their relevance. Our research supports the assumption of different faking processes and explains why detecting faking is a complex endeavor.