AI Can Create More “Real” Faces Than Real Humans, Study Finds
A recent study in the journal Sage has revealed that artificial intelligence (AI) is capable of creating faces that appear more “real” to humans than actual photographs of real human faces. This phenomenon, described as “hyperrealism,” poses a concerning implication in relation to the growing prevalence of deepfakes – artificially generated material created to impersonate real individuals.
However, a crucial finding from the study is that AI achieved hyperrealism exclusively when generating white faces, as AI-generated faces of color failed to reach the same effect, falling into what is known as the “uncanny valley.” Amy Dawel, the senior author of the study and a clinical psychologist and lecturer at the Australian National University, emphasized that this racial bias could have significant implications not only for the construction of AI tools, but also for how people of color are perceived online.
The study was inspired by previous research published in the journal PNAS in February 2022, which found that participants were unable to distinguish between AI-generated and human faces. Dawel and her team wanted to assess whether there was a racial component to how people perceive AI-generated faces, prompting the new study that specifically focused on white participants.
The findings indicate that AI algorithms, such as the StyleGAN2 image-generation tool, tend to be disproportionately trained on white faces. This training bias contributes to the hyperreal effect seen in white faces generated by AI. Additionally, Dawel highlighted examples of racial bias in AI systems, such as the alteration of skin tone and eye color for people of color in tools used to create professional headshots from regular photos.
These biases could have significant social consequences, as AI-generated images are increasingly utilized in areas such as marketing, advertising, and illustration. The use of AI, if developed with biases, may perpetuate racial prejudices in the media, leading to deep-seated consequences on a societal level.
One of the most concerning findings from the study is that individuals who were most confident in their choices were also the ones who made the most mistakes in identifying AI-generated faces as real. This lack of awareness among those who are most easily fooled by AI further underscores the need for transparency and monitoring of generative AI.
Dawel argued that transparency and independent oversight are essential in the development and monitoring of AI technology. She emphasized the importance of directing resources and funding towards research on AI with a focus on minimizing biases and mitigating risks associated with the technology’s development.
Ultimately, the study calls for increased regulation and oversight to address the biases in AI development, as well as for companies creating AI to allocate resources towards independent research efforts. With the rapid advancement of AI technology, the need for action to ensure the responsible and ethical development of AI has become more urgent than ever.