Emotion Recognition From AI-Generated and Human Faces: Performance, Mental Health, and Demographic Diversity
Abstract
Facial emotion-recognition tasks are widely used to investigate social-cognitive differences across mental health conditions, but their scientific value depends on the faces used to measure performance. Established photographic databases are difficult and expensive to expand, and many provide little ethnic diversity. AI-generated faces offer a more representative, flexible, and scalable alternative, provided they capture the same clinically relevant individual differences as photographs of human faces. We compared 161 photorealistic AI-generated faces with 72 images from FACES, an established human-photograph set. Prolific participants contributed 83,206 valid trials; 77,238 trials covering the six shared emotions entered the primary comparison (accuracy \textit{n} = 1,047; response time \textit{n} = 1,046). AI-diverse expressions were recognised more accurately (89.12\% vs.\ 81.72\%) and faster (2.29 vs.\ 2.64 s) than FACES expressions, while AI-diverse and AI-White performance did not differ significantly. The AI task retained the individual differences needed for research: lower accuracy on both sources was associated most consistently with psychotic-like experiences and obsessive-compulsive symptoms, with magnitudes aligned with previous literature. AI-diverse and FACES accuracy correlated at \textit{r} = .56, and corrected split-half reliability was .70 and .60, respectively. Women were more accurate than men; participant ethnicity was unrelated to performance. Higher AI accuracy indicates that the expressions were clear; it does not show general psychometric superiority, possibly because very easy items can compress differences among high performers. The combination of clear expressions, comparable symptom sensitivity, and broader representation shows that generative AI can modernise a core social-cognition research tool while making stimulus design more controllable, expandable, and shareable.
Related
- Performance and response time in emotion recognition and mentalizing measures in patients with schizophrenia.
- The Relationship Between Childhood Maltreatment and Emotion Recognition in Adults: A Systematic Review and Meta-Analysis
- Emotion Recognition Impairments in People with Dementia are Associated with Lower Physiological Linkage with Caregivers
- A robust adversarial learning framework for EEG emotion recognition
- Facial emotion recognition of sadness relates to superior parietal gray matter volume in mild cognitive impairment.
BibTeX
@misc{vowels2026emotion,
title = {{Emotion Recognition From AI-Generated and Human Faces: Performance, Mental Health, and Demographic Diversity}},
author = {Matthew J Vowels and Laura M Vowels},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/gp2jq_v2},
url = {https://osf.io/preprints/psyarxiv/gp2jq_v2/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Emotion Recognition From AI-Generated and Human Faces: Performance, Mental Health, and Demographic Diversity
AU - Matthew J Vowels
AU - Laura M Vowels
PY - 2026
DA - 2026/09/19
PB - PsyArXiv
DO - 10.31234/osf.io/gp2jq_v2
UR - https://osf.io/preprints/psyarxiv/gp2jq_v2/
AB - Facial emotion-recognition tasks are widely used to investigate social-cognitive differences across mental health conditions, but their scientific value depends on the faces used to measure performance. Established photographic databases are difficult and expensive to expand, and many provide little ethnic diversity. AI-generated faces offer a more representative, flexible, and scalable alternative, provided they capture the same clinically relevant individual differences as photographs of human faces. We compared 161 photorealistic AI-generated faces with 72 images from FACES, an established human-photograph set. Prolific participants contributed 83,206 valid trials; 77,238 trials covering the six shared emotions entered the primary comparison (accuracy \textit{n} = 1,047; response time \textit{n} = 1,046). AI-diverse expressions were recognised more accurately (89.12\% vs.\ 81.72\%) and faster (2.29 vs.\ 2.64 s) than FACES expressions, while AI-diverse and AI-White performance did not differ significantly. The AI task retained the individual differences needed for research: lower accuracy on both sources was associated most consistently with psychotic-like experiences and obsessive-compulsive symptoms, with magnitudes aligned with previous literature. AI-diverse and FACES accuracy correlated at \textit{r} = .56, and corrected split-half reliability was .70 and .60, respectively. Women were more accurate than men; participant ethnicity was unrelated to performance. Higher AI accuracy indicates that the expressions were clear; it does not show general psychometric superiority, possibly because very easy items can compress differences among high performers. The combination of clear expressions, comparable symptom sensitivity, and broader representation shows that generative AI can modernise a core social-cognition research tool while making stimulus design more controllable, expandable, and shareable.
ER -