Even when specifically instructed to do otherwise, the AI tool exhibited a propensity to produce images of people that were the same race.
According to a study by The Verge, Meta AI is regularly failing to generate correct photos for seemingly straightforward cues such as “Asian man and Caucasian friend” or “Asian man and white wife.” The company’s picture generator, on the other hand, appears to have a preference for producing images of people of the same race, even when it is expressly told to produce images of people of different races.
These findings were validated by Engadget through our own use of Meta’s web-based image creator for testing purposes. Images of Asian couples were generated when the prompts “an Asian man with a white woman friend” or “an Asian man with a white wife” were used. In response to the request for “a diverse group of people,” Meta AI produced a grid consisting of nine white faces and one person of color. There were a few instances in which it produced a single result that accurately mirrored the prompt; nevertheless, in the majority of instances, it completely failed to accurately describe the prompt.
As The Verge points out, there are other more “subtle” evidence of prejudice in Meta AI, such as a tendency to make Asian males appear older while Asian women appeared younger. This is one example of the bias within Meta AI. When the prompt did not include the phrase “culturally specific attire,” the picture generator might occasionally add it to the list of possible responses.
It is not quite obvious why Meta AI is having difficulty with these kinds of requests; nevertheless, it is not the only generative AI platform to come under attention for the way it portrays racial issues. After overcorrecting for diversity and producing weird results in response to cues regarding historical individuals, Google’s Gemini image generator temporarily halted its capacity to create images of humans. After some time had passed, Google provided an explanation that its internal safeguards had failed to take into account instances in which diverse results were improper.
In response to a request for comment, Meta did not immediately provide a response. Prior to this, the company has stated that Meta AI is now in the “beta” stage, which means that it is prone to making errors. It has also been difficult for Meta AI to provide accurate responses to straightforward questions concerning prominent people and current events.