Some AI-generated visages can now be so realistic that they are indistinguishable from actual human faces in photographs. This revelation stems from a recent study published in the book Psychological Science.
Conducted by experts from universities in Canada, Australia, and London, the study comprised two research experiments aimed at evaluating participants’ ability to discern AI-generated content.
In one of the studies, where 124 participants were tasked with differentiating between AI-created and authentic faces, American AI faces were more frequently perceived as genuine compared to actual human faces. Participants were also required to assess their confidence levels on a 100-point scale while making their judgments.
The results indicated that while 51% of real images were correctly identified as authentic, a higher percentage of AI-generated images, particularly those featuring pale faces, were mistakenly categorized as real, with a 66% accuracy rate.
The study’s researchers highlighted a startling observation, noting that white AI-generated faces could appear more authentic than real human faces, deceiving observers without their awareness.
Interestingly, the outcomes varied for images portraying individuals of color. The report suggested that the prevalence of pale faces in the AI training data contributed to their exceptionally realistic depiction.
Zak Witkower, one of the study’s authors, emphasized that AI algorithms tend to excel in generating realistic depictions of white faces compared to faces of other ethnicities.
The research team cautioned that such biases could lead to tangible consequences if left unaddressed, potentially influencing critical areas such as elections or search efforts for missing persons.
Evidence of these biases can be traced back to the training data and outcomes of prominent AI models, including early versions of ChatGPT. Researchers have long warned about the ramifications of unchecked biases in AI models, particularly the adverse impact they could have on marginalized communities with the widespread integration of such technologies.
