EEG EVIDENCE FOR DIFFERENTIATING DEEPFAKE AND HUMAN SPEAKERS

Deepfake videos increasingly resemble real humans, raising questions about how the brain distinguishes artificial from human communicators during emotional narrative processing. This study investigated how real-human versus AI-generated video stories modulate EEG oscillations and event-related potentials (ERPs). Thirty adults viewed short everyday-life narratives presented either by a human speaker or a deepfake matched in appearance and motion, with stories varying in emotional valence. EEG time–frequency power (theta, alpha, beta) was extracted from frontal, parietal, temporal, and occipital regions, and ERPs (P1, N170, EPN, LPP) were analyzed with region-specific tests. Human speakers elicited stronger frontal beta and theta power than deepfakes, whereas deepfake videos produced greater alpha, beta, and theta power in parietal and occipital regions, revealing a clear anterior–posterior dissociation. Emotional valence effects were agent-dependent, with frontal alpha being higher for negative than for positive stories across agents, while frontal theta encoded valence only for human speakers. In contrast, parietal oscillations reflected valence exclusively for deepfakes, showing increased theta for negative narratives. ERP results showed early discrimination of agents. P1 and N170 amplitudes differed between human and deepfake videos depending on region and emotional condition. Later components primarily encoded emotion: LPP differentiated positive from negative narratives, and EPN showed a valence effect only for deepfake stimuli in frontal regions. Overall, deepfake communicators evoked distinct temporal and spatial neurocognitive signatures compared with real humans, highlighting differential processing of speaker identity and emotional meaning across early perceptual and later evaluative stages.