Images of faces generated by artificial intelligence (AI) can now appear more real than pictures of real people—as long as they are white, a study said on Monday.
But AI tends to depict people of other ethnicities with white features because the data used to train the algorithms is biased, said lead author Amy Dawel from the Australian National University (ANU).
“If white AI faces are consistently perceived as more realistic, this technology could have serious implications for people of color by ultimately reinforcing racial biases online,” she said.
The authors said it was vital to educate the public about the hyper-realism of today’s AI to avoid misinformation and online scams.
The researchers used 200 pictures—100 AI-generated and 100 real humans—and asked 124 participants which were AI and which were genuine.
Of the five faces most often identified as human, four were AI. Of the top five judged as AI, four were human.
The researchers ran another experiment to try to understand why this was the case, asking more than 600 people to rate various qualities of the images—attractiveness, familiarity, image quality.
Clear differences remained between AI and real images, said the researchers, but the participants were either misinterpreting the details or missing them entirely.
For example, participants tended to judge faces as human if they were more proportional, less memorable and more familiar—but these qualities were more prominent in the AI images.
The researchers used the data based on these perceptions to program a model that was able to accurately identify more than 90 percent of the faces.
“However, we can’t rely on these physical cues for long,” said Dawael.
“AI technology is advancing so quickly that the differences between AI and human faces will probably disappear soon.”
AI Hyper-realism: Why AI Faces Are Perceived As More Real Than Human Ones, Psychological Science (2023). DOI: 10.1177/09567976231207095
© 2023 AFP
AI images of white faces are now ‘hyper-real’: Study (2023, November 13)
retrieved 13 November 2023
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no
part may be reproduced without the written permission. The content is provided for information purposes only.