Meta’s artificial intelligence-powered text-to-image generator refuses to render photos of an Asian person dating a white man or woman — even though the company’s billionaire founder Mark Zuckerberg is married to the daughter of Chinese immigrants.
The AI tool called Imagine was released last December and seems to be suffering from a similar revisionist malady that afflicted Google’s AI image generator Gemini, which was paused after outcry over it spitting out wildly inaccurate depictions, including Native American popes and black Founding Fathers.
On Friday, The Post urged Imagine to create photos with prompts that read “Asian man and Caucasian friend,” “Asian man and white wife,” and “Asian woman and Caucasian husband.” All of the queries returned results that only showed two Asian people.
However, when the software was asked to create an image of an Asian woman with a black friend, it generated an accurate result. It did the same when asked to create romantic pairings between white and black people.
Critics have long contended that the algorithms used by AI image generators reflect the biases of the software engineers who program them.
Imagine’s refusal to render a white person dating an Asian person, first reported by The Verge, is head-scratching considering that Zuckerberg is married to Priscilla Chan. They have three kids.
The Post reached out to Meta for comment.
The image generator comes with a disclaimer that warns some of its photos “may be inaccurate or inappropriate.”
Before Imagine was launched, Meta posted a blog item saying that its generative AI software was being programmed in a way that would “reduce bias.”
“Addressing potential bias in generative AI systems is a new area of research,” the blog post, written last September, states.
“As with other AI models, having more people use the features and share feedback can help us refine our approach.”
Imagine’s shocking results come as the social media giant on Friday announced major changes to its policies on digitally created and altered media ahead of elections to test its ability to police deceptive content generated by new artificial intelligence technologies.
Meta — which owns Facebook, Instagram WhatsApp, and Threads — will start applying “Made with AI” labels in May to AI-generated videos, images, and audio posted on its platforms, expanding a policy that previously addressed only a narrow slice of doctored videos, Vice President of Content Policy Monika Bickert said in a blog post.
Google said last year that AI labels are coming to YouTube and its other platforms.
With Post Wires
Source
