Does ChatGPT always create the impression that you are interacting with another individual?
Artificial intelligence (AI) has progressed to such an extent that some tools can convincingly mimic human interaction, leading to a remarkable level of realism. This advancement extends beyond mere conversation. A recent study published in Psychological Science revealed that images of pale faces generated by the renowned StyleGAN2 engine can sometimes appear more “human” than actual human eyes.
The Surreal Eyes Crafted by AI
To investigate this phenomenon, we conducted a study involving 124 participants who were shown a series of images featuring various pale faces. Their task was to differentiate between AI-generated faces and real faces.
Half of the images depicted genuine faces, while the other half were AI-generated. If participants had chosen randomly, they would have been correct approximately half of the time, akin to a coin flip.
However, the results were surprising. Participants consistently mistook AI-generated faces for real ones. In fact, nearly two-thirds of the AI-generated faces were misclassified as human by the participants.
These findings suggest that AI-generated faces exhibit a heightened sense of realism, a concept termed “hyperrealism.” Moreover, they indicate that people often struggle to discern AI-generated faces from real ones. You can compare the real-life photos at the top of the page with the AI-generated ones above to see for yourself.
But what if individuals are cognizant of their limitations and are less likely to encounter AI-generated faces online?
To delve deeper, we assessed the participants’ confidence levels in their decisions. Surprisingly, those who had the most difficulty identifying AI faces were the most confident in their judgments.
In essence, those deceived by AI were unaware of being deceived.
Biased Outcomes Stemming from Distorted Training Data
The advent of the fourth industrial revolution has reshaped the online landscape, introducing solutions like AI, automation, and advanced technologies, including AI-generated faces.
While AI-generated faces offer various benefits, they also pose risks. They have been implicated in identity theft, phishing scams, and cyber warfare, in addition to aiding in locating missing individuals.
Due to overconfidence in their ability to distinguish AI faces, individuals may be more susceptible to deceptive practices. For instance, they might unknowingly divulge sensitive information to cybercriminals posing as realistic AI personas.
Another troubling aspect is the ethnic bias present in AI hyperrealism. Our research, drawing from a separate study that examined Asian and Black faces, revealed that only light-skinned AI-generated faces were perceived as realistic.
When participants were asked to identify whether the faces were human or AI-generated, they could accurately identify faces of color only half of the time, akin to random guessing.
This bias underscores the fact that white AI-generated faces are perceived as more realistic than actual white human faces or artificially altered faces.
Implications of Bias and Lifelike AI
This cultural bias likely stems from the predominant use of images of predominantly white faces in training AI models.
Racial bias in computational training can have serious consequences. For instance, self-driving vehicles are less adept at detecting Black individuals, putting them at greater risk than their white counterparts. It is incumbent upon both AI developers and regulatory bodies to ensure diverse representation and mitigate bias in AI technologies.
The proliferation of lifelike AI also raises concerns about our ability to discern genuine content and protect ourselves against deception.
Our research identified specific characteristics that lend a realistic appearance to pale AI faces. These faces often possess generic features without distinctive traits that would set them apart as “unique.” Participants mistook these features for indicators of human authenticity, contributing to the hyperrealism effect.
Given the rapid pace of AI advancement, it will be intriguing to see how these findings hold up over time. Furthermore, there is no guarantee that AI-generated faces from different algorithms will mirror human faces in every aspect.
Subsequent to our study, we evaluated AI surveillance technology’s capability to identify the AI faces we used. Despite the technology’s claims of high precision, its performance was comparable to that of individual participants—subpar.
Similarly, AI writing tools have been known to erroneously flag non-English speakers, particularly those using tribal languages, for plagiarism.
Mitigating the Challenges Posed by AI
How can individuals safeguard themselves against mistaking AI-generated content for authentic material?
One approach is to acknowledge the challenges people face in distinguishing between AI-generated and human content. By recognizing our limitations in this regard, we may become more discerning consumers of online content and take steps to verify information when necessary.
Public policy plays a crucial role in addressing this issue. One option is to mandate the disclosure of AI usage. This transparency can prevent misuse of AI for deceptive purposes and foster a sense of trustworthiness in online interactions.
Another strategy involves prioritizing verification of credible sources. Implementing a trusted source badge, earned through rigorous authentication processes, can assist users in identifying reliable media sources, akin to labels like “Made in Australia” or the “European CE tag.” 
