- AI-generated white faces are more convincing than photographs of real humans, a study has found.
- However, the results were not the same for AI images of people of color.
- This may be because most AI models are disproportionately trained on white faces, the study found.
Some AI-generated faces are already indistinguishable from photographs of real humans.
At least, that’s what a new study published in the Psychological Science journal found.
The study, which was conducted by researchers from universities in Canada, Australia, and London, carried out two experiments that tested people’s ability to identify AI-created content.
In one, where 124 participants were asked to judge whether a face was AI-generated or real, the researchers found that caucasian AI faces were judged as human more often than photographs of real human faces. Participants were also asked to rate their confidence in their answers on a 100-point scale.
The study found that 66% of AI images of white faces were rated as human, while 51% of real images were identified as such, the study found.
“Remarkably, white AI faces can convincingly pass as more real than human faces — and people do not realize they are being fooled,” the researchers said in the study.
However, the results were different for pictures of people of color. The report noted that this may be because most AI models were disproportionately trained on white faces, which made them appear “especially realistic.”
One of the study’s authors, Zak Witkower, told The Guardian that this meant AI was “going to produce more realistic situations for white faces than other race faces.”
The team warned that if the issues remained unaddressed, such bias could have real-world consequences, ranging from “influencing elections to finding missing children.”
There have been several examples of instances of basis found in the training data and output of popular AI models, including early versions of ChatGPT.
Researchers have warned for some time that the widespread adoption of AI models could cause trouble for marginalized groups if the issues are not properly addressed.