Classifying AI-Generated and Real Faces from EEG: A Comparative Evaluation of LDA, Logistic Regression, and a Spatiotemporal CNN

Hyperrealistic artificial intelligence (AI) generated faces are often misclassified as real, although spontaneous neural activity may still encode their artificial origin. We performed a secondary multivariate reanalysis of electroencephalography (EEG) data acquired from 26 participants viewing real and StyleGAN2-generated faces during an orthogonal target-detection task. Condition-averaged Event-Related Potentials (ERPs) from 14 selected channels and four regional Principal Component Analysis (PCA) features within 200–600 ms were analyzed using mass-univariate mapping, ridge encoding, group-aware logistic decoding, classical machine learning, a spatiotemporal Convolutional Neural Network (STCNN), and explainable-AI methods. Linear encoding was weak, whereas logistic decoding reached 0.623 ± 0.106 accuracy. In a 25-fold classical machine-learning benchmark, Linear Discriminant Analysis achieved 0.700 accuracy, increasing to 0.720 after optimization (Receiver Operating Characteristic Area Under the Curve (ROC-AUC) = 0.780), with permutation testing confirming above-chance classification (p = .001). The optimized STCNN achieved 64.0% ± 5.8% accuracy and AUC = 0.670, with significant held-out classification performance (p = .0028). Explainability analyses revealed a predominantly right-lateralized posterior occipito-temporal/temporo-parietal pattern with frontal involvement, with saliency peaks around 298, 355, and 500 ms. These findings provide preliminary evidence that late, distributed ERP activity contains decodable information distinguishing AI-generated from real faces.