# Uomo o Macchina: il cervello conosce la differenza? | Giulia Panozzo | TEDxVicenza Source: https://www.youtube.com/watch?v=JElJ0C66t7Q Recap page: https://rapidrecap.app/video/JElJ0C66t7Q Generated: 2025-12-19T19:36:03.821+00:00 --- ## Quick Overview The brain can still distinguish between human and hyper-realistic AI-generated faces, although accuracy drops to 78% for hyper-realistic static faces and reaction times slow down, demonstrating that despite technological progress, our instinctive response mechanisms still detect subtle differences, especially when movement and sound are introduced. **Key Points:** - Participants incorrectly classified hyper-realistic AI faces as human 22% of the time (accuracy 78%) compared to 10% for non-realistic AI faces (accuracy 90%). - Reaction times were significantly slower for both AI conditions (1473.48 ms for hyper-realistic vs. 1519.89 ms for real humans), indicating cognitive effort. - The experiment involved showing participants static images of humans, hyper-realistic AI, and 'perturbing' AI, measuring accuracy and reaction time. - The 'perturbing' AI stimuli (which included strange proportions, colors, or movements) resulted in 99% accuracy, showing that our instincts quickly identify non-human cues. - The speaker argues that the challenge is not the AI itself, but how we regulate its use in areas like advertising and social interaction, given the difficulty in distinguishing reality from simulation. - When participants viewed video stimuli, the difference in emotional response (measured via skin conductance) was greater for human stimuli than for highly realistic AI stimuli, suggesting a persistent physiological difference. ![Screenshot at 01:40: The slide displays the Uncanny Valley model, illustrating that as human likeness increases \(x-axis\), familiarity \(y-axis\) generally rises, but dips sharply into a negative zone \('valle perturbante'\) for near-human stimuli like zombies or corpses, before rising sharply again for healthy humans.](https://ss.rapidrecap.app/screens/JElJ0C66t7Q/00-01-40.png) **Context:** Giulia Panozzo discusses the psychological and physiological responses to increasingly realistic artificial intelligence-generated faces, focusing on the phenomenon known as the 'Uncanny Valley' (or 'valle perturbante' in Italian). The presentation references research from Professor Masairi Mori and subsequent studies to explore whether the human brain's instinctive response mechanisms can reliably differentiate between real humans and highly realistic AI creations, especially in contexts like marketing and social interaction. ## Detailed Analysis Giulia Panozzo addresses the challenge of distinguishing between human and artificial entities as AI generation technology advances. She introduces Mori's 'Uncanny Valley' model, which maps familiarity against human likeness, showing that entities that are *almost* human evoke discomfort. Panozzo highlights that even in the 1970s, researchers observed that human brains reacted negatively to stimuli that were too close to human but imperfect. Her own research confirmed this effect using static images: participants correctly identified real humans 90% of the time, hyper-realistic AI faces 78% of the time, and 'perturbing' AI (with strange features) 99% of the time. Crucially, reaction times were slowest for the near-human AI, suggesting the brain expends more effort processing them. Furthermore, physiological stress responses (skin conductance) were significantly higher when viewing human stimuli compared to highly realistic AI stimuli, even when participants misidentified the AI as human. This implies that our nervous system retains an instinctive ability to detect subtle differences, even when our conscious judgment fails. Panozzo concludes by emphasizing the need for ethical regulation and conscious decision-making regarding the integration of such advanced AI into daily life, particularly in persuasive contexts like advertising. ### The Uncanny Valley - The concept posits that familiarity with non-human entities drops sharply when they achieve near-human likeness (e.g., corpses, zombies), even if hyper-realistic AI is improving. ### Experimental Results - Accuracy in identifying real humans was 90%, dropping to 78% for hyper-realistic AI faces, while reaction time was slowest for AI stimuli, indicating cognitive strain. ### Perturbing Stimuli - AI stimuli featuring unnatural proportions or movements were easily identified as non-human (99% accuracy), confirming the brain's reliance on specific cues. ### Physiological Response - Skin conductance measurements showed a significantly stronger stress response to real human stimuli compared to highly realistic AI stimuli, suggesting an underlying biological detection mechanism. ### Implications for AI Use - The speaker stresses the ethical responsibility to regulate AI use in marketing and persuasion, as our subconscious may be reacting to these highly realistic stimuli differently than to actual humans, leading to potential manipulation. ### Conclusion - The gap between human experience and AI simulation is still detectable by our primal nervous system, necessitating conscious navigation of this new era of human-machine interaction. ![Screenshot at 00:03: Initial slide displaying institutional and visionary partners of the event, including the City of Vicenza and H-Farm International School.](https://ss.rapidrecap.app/screens/JElJ0C66t7Q/00-00-03.png) ![Screenshot at 00:10: Speaker Giulia Panozzo on stage at TEDxVicenza, with a screen showing two faces \(Asian female and Middle Eastern male\) used in the experiment setup.](https://ss.rapidrecap.app/screens/JElJ0C66t7Q/00-00-10.png) ![Screenshot at 01:40: Detailed graph illustrating the Uncanny Valley, plotting 'familiarity' against 'human likeness' for moving and still objects, showing the dip corresponding to the 'uncanny valley'.](https://ss.rapidrecap.app/screens/JElJ0C66t7Q/00-01-40.png) ![Screenshot at 03:33: Slide showing the results table: Human \(90% accuracy, 1519.89ms RT\), Hyper-realistic AI \(78% accuracy, 1473.48ms RT\), Perturbing AI \(99% accuracy, 925.86ms RT\).](https://ss.rapidrecap.app/screens/JElJ0C66t7Q/00-03-33.png) ![Screenshot at 06:29: Diagram illustrating the experimental sequence involving static images of human and AI faces with varying exposure times \(60s, 30s, repeated 2x\).](https://ss.rapidrecap.app/screens/JElJ0C66t7Q/00-06-29.png)