Presentation
Understanding Human Perception and Reasoning towards Human-AI teaming in Deepfake Detection
DescriptionDeepfakes are a form of artificial intelligence (AI)-generated synthetic media that can manipulate the lip movements, facial expressions, and voices of specific individuals. This technology poses serious risks to individuals and society by enabling more convincing deception, misinformation, and identity misuse. Although AI-based detection tools are essential for detecting Deepfake, understanding and strengthening human detection ability is equally important.
In this paper, we compare the attention allocation patterns of humans and AI systems in Deepfake detection, showing that they rely on distinct cues and reasoning strategies. We further examine physiological differences among individuals with varying levels of AI literacy to identify factors associated with superior human performance. Contrary to the assumption that AI systems consistently outperform humans in pattern recognition and classification, our results suggest that people with greater prior exposure to AI can rival or even exceed AI-based detectors in certain Deepfake contexts. Eye-tracking and EEG evidence reveal humans’ distinctive capacity for holistic, intuitive, and flexible visual-cognitive processing, which may enable them to detect subtle inconsistencies and artifacts overlooked by AI.
This study advances understanding of how humans and AI process deceptive visual cues differently and informs the design of trustworthy and adaptive human-AI teaming for Deepfake detection.
In this paper, we compare the attention allocation patterns of humans and AI systems in Deepfake detection, showing that they rely on distinct cues and reasoning strategies. We further examine physiological differences among individuals with varying levels of AI literacy to identify factors associated with superior human performance. Contrary to the assumption that AI systems consistently outperform humans in pattern recognition and classification, our results suggest that people with greater prior exposure to AI can rival or even exceed AI-based detectors in certain Deepfake contexts. Eye-tracking and EEG evidence reveal humans’ distinctive capacity for holistic, intuitive, and flexible visual-cognitive processing, which may enable them to detect subtle inconsistencies and artifacts overlooked by AI.
This study advances understanding of how humans and AI process deceptive visual cues differently and informs the design of trustworthy and adaptive human-AI teaming for Deepfake detection.
Event Type
Lecture
TimeTuesday, October 20th3:40pm - 4pm PDT
Location