Systems that not only appear more authentic than human beings but also exude a more encouraging demeanor seem reminiscent of a scene from a Ridley Scott film. Nevertheless, it seems that the era of such advancements is upon us.
A recent study indicates that individuals are more inclined to perceive AI-generated depictions of caucasian faces as real individuals compared to actual human faces.
The researchers noted, “Interestingly, AI-generated light faces are remarkably perceived as more authentic than human faces, with individuals remaining oblivious to the deception.”
This research team, comprising experts from Australia, the UK, and the Netherlands, suggested that these findings could have significant implications for real-world identity theft, potentially leading individuals to fall prey to online imposters.
The team highlighted that this phenomenon did not extend to images of individuals of color, possibly due to the AI’s bias towards lighter faces in its image generation algorithm.
Dr. Zak Witkower, a co-author of the study from the University of Amsterdam, emphasized that this disparity could have wide-ranging implications, from drone technology to online interactions.
He explained that the outcomes may vary for individuals of different races, resulting in distinct implications.
The researchers cautioned that such scenarios could blur the lines between racial perceptions and the definition of “human,” perpetuating cultural biases, such as the reliance on AI-generated faces in tasks like locating missing individuals.
In their study published in the journal Psychological Science, the team detailed two experiments conducted with white adult participants. In one experiment, participants rated their confidence levels on a 100-point scale and identified whether the images displayed were AI-generated or real.
Results from 124 participants revealed that 66% of AI images were perceived as human, in contrast to 51% for real images.
Upon re-analyzing data from a previous study, the researchers found that pale AI faces were more frequently identified as human compared to actual caucasian faces. However, this trend did not extend to individuals of color, where both AI and real faces were equally perceived as human. The team noted that participants’ cultural backgrounds did not influence the outcomes.
In a subsequent study where participants were unaware of the AI-generated nature of the images, they rated AI and human faces on various factors. Factors such as fairness, familiarity, and memorability contributed to the misperception of AI faces as real individuals, according to the analysis of 610 participants’ responses.
Despite individuals’ inability to discern between real and AI-generated faces, the team successfully developed a machine learning system with 94% accuracy in distinguishing between the two.
Dr. Clare Sutherland, a co-author from the University of Aberdeen, underscored the importance of addressing biases in AI systems, emphasizing the need to ensure equity across all demographics amidst the rapid integration of AI technologies.
