Measuring Emotion Perception Ability Using AI-Generated Stimuli: Development and Validation of the PAGE Test

Abstract

We present a new measure of emotion perception called PAGE (Perceiving AI Generated Emotions). The test includes 20 emotions, expressed by ethnically diverse faces, spanning a wide range of ages. We created stimuli with generative AI, illustrating a method to build customizable assessments of emotional intelligence at relatively low cost. Study 1 describes the validation of the image set and test construction. Study 2 reports the psychometric properties of the test, including convergent validity and relatively strong reliability. Study 3 explores predictive validity using a lab experiment in which we causally identify the contributions managers make to teams. PAGE scores predict managers' causal contributions to group success, a finding which is robust to controlling for personality and demographic characteristics. We discuss the potential of generative AI to automate development of non-cognitive skill assessments.

Keywords: emotion perception; generative AI; management; measurement; teamwork.

Figure A1

Study 2a: Scree plot of PAGE. The plot displays the eigenvalues for the components extracted with principal component analysis (PCA). The sharp drop between the first and second components indicates a dominant one-factor structure.

Figure 1

Example Stimuli from the PAGE test. In this case, the emotions being represented by the four stimuli are (a) surprise, (b) contentment, (c) anger, (d) fear.

Figure 2

Sample item from the PAGE test. For each item, participants select one answer from six options. Definitions of emotions are provided to participants if they click on the question marks. The target emotion for this item is fear.

Figure 3

Distribution of PAGE scores by gender (N = 505 males and 505 females). Mean score for male is 23.4 (SD = 4.9), Mean score for female is 24.0 (SD = 5.1), mean scores indicated by the vertical lines. The mean difference between males and females is not statistically significant.

Figure 4

Patterns of performance for PAGE/RMET by gender/age/ethnicity (N = 741). 50% of the participants are female, and 50.5% are White. Both tests show similar performance patterns across gender, age and ethnicity. Each dot on the graphs represents one observation.

Figure 5

Experiment overview. Participants were randomly assigned to the role of ‘manager’ or ‘worker’. Each group completed a novel collaborative problem-solving task in which managers assigned tasks, monitored group progress, and motivated teammates. After each task, managers were randomly reassigned to new groups, managing a total of four different groups. The repeated random assignment of managers to teams enabled the identification of each manager’s contribution to group performance.