In a scenario reminiscent of a sci-fi plot, technology has reached a point where AI-generated faces not only sound more authentic than actual humans but also appear more convincing. A recent study has revealed a striking tendency among individuals to mistake AI-generated images of white faces for real humans, surpassing the credibility of photographs of actual individuals.
The research, conducted by a team comprising experts from Australia, the UK, and the Netherlands, underscores the potential implications of this phenomenon in real-world scenarios, particularly in the realm of identity theft. The researchers note the possibility of individuals unknowingly falling victim to digital impostors.
Interestingly, the study’s results did not extend to images of people of colour, a nuance that the researchers attribute to the predominantly white-focused training data used to develop the AI face-generating algorithm.
Dr. Zak Witkower, a co-author from the University of Amsterdam, emphasizes the broader consequences of this bias, impacting areas such as online therapy and robotics. He highlights the likelihood of AI creating more realistic situations for white faces compared to other racial groups.
The researchers caution that this situation could intertwine perceptions of race with perceptions of humanity, potentially perpetuating social biases. The implications extend to critical areas such as finding missing children, where reliance on AI-generated faces may introduce biases.
Outlined in the journal Psychological Science, the study comprised two experiments. In one, white adults assessed images of AI-generated and real white faces, with results indicating that 66% of AI images were perceived as human compared to 51% of real images. The second experiment involved participants rating AI and human faces on various attributes without knowing some were AI-generated, revealing factors such as facial proportionality, familiarity, and memorability as key contributors to the mistaken belief that AI faces were human.
Ironically, even as humans struggle to distinguish AI-generated faces from real ones, the researchers developed a machine learning system capable of achieving 94% accuracy in this task.
Co-author Dr. Clare Sutherland from the University of Aberdeen emphasizes the study’s call to address biases in AI, underscoring the need to ensure that the rapid integration of AI does not leave any individual or group disadvantaged based on factors such as ethnicity, gender, or age.