Are we sure that we can distinguish a real face from an artificially generated one? Today, artificial image-generation systems produce faces so convincing that it is almost impossible to distinguish the real from the artificial. In an experiment conducted at the Cognitive Electrophysiology Laboratory of the Department of Psychology at the University of Milano-Bicocca, a group of students was asked to determine whether the faces presented were real or generated by artificial intelligence and how attractive or familiar they appeared. Artificial faces were correctly identified in only 33 per cent of cases, well below chance level, which is 50 per cent. Yet the brain, monitored in real time using high-density electroencephalography with 128 channels, was not deceived, as shown by electrical potentials and by the analysis of intracortical sources through low-resolution electromagnetic tomography, which makes it possible to assess the activity of different areas of the brain.
The research, conducted by Professor Alice Mado Proverbio and Dr Mariia Dosaikina, demonstrates a profound dissociation between conscious judgement and the implicit neural response to AI-generated faces: the human brain processes them in a significantly different way from real faces and continues to distinguish them even when the individual is unable to recognise them as artificial. The study, funded by the University and the Italian Ministry of Universities and Research, was published in Scientific Reports under the title “Neural signatures of hyper-realistic AI-generated faces: dissociating behavioral indistinguishability from implicit neural evaluation” and shows that faces generated by artificial intelligence have crossed the threshold of the “uncanny valley”.
The concept of the “uncanny valley” describes the psychological discomfort and sense of unease that we experience when faced with human figures that are not entirely convincing, such as robots, mannequins and 3D animations. Current GAN technologies — Generative Adversarial Networks — have, however, made a leap forward: the most advanced synthetic faces appear to have crossed the uncanny valley at the behavioural level, thanks to the paradox of faces that are too perfect to cause discomfort. On the contrary, participants in the study judged them to be more attractive, familiar and trustworthy than real human faces. This familiarity does not derive from actual knowledge of the faces, but from the structural characteristics introduced by generative algorithms, which tend to create more prototypical faces. Unlike human faces, shaped by genetic mechanisms that favour the uniqueness of individuals, synthetic faces appear to evoke an intrinsic sense of familiarity through the optimisation of common and typical features, which the brain recognises as more familiar and, consequently, more attractive.
«The electrophysiological data,» explains Professor Proverbio, «show that the typical rejection response associated with the “uncanny valley” has not disappeared: it has shifted to a deeper, neural and unconscious level. Even though the conscious mind no longer recognises it, the brain still perceives it through an accurate unconscious neural assessment.»
Thirty different participants viewed 440 images of faces, half real and half artificial, while their brain activity was recorded. The participants were completely unaware of the presence of AI-generated faces and were focused on detecting the presence of fictitious visual targets. The results show that artificial faces, in addition to eliciting greater familiarity and aesthetic appreciation, more strongly activate the ventro-temporal, parietal, limbic and medial frontal brain networks, associated with affective evaluation, the attribution of a mind to social agents and recognition memory. The brain also recognises them more quickly, as though they were clearer and more prototypical stimuli.
«We have entered a new era. These findings,» adds Professor Proverbio, «could have significant implications for digital security, communication and social media. If AI-generated faces are perceived as more reassuring than real ones, this represents a systemic cognitive vulnerability: people tend to trust more what they perceive as familiar. Deepfakes do not merely deceive the “eye”; they act directly on the neural mechanisms of trust and recognition.»