From The Atlantic
Hot AI Jesus Is Huge on Facebook
By Caroline Mimbs Nyce
Jesus is punching the devil on Facebook.
The two are in a boxing ring. Jesus is wearing a pair of white boxing shorts with his name embroidered on the waistband. He is ripped beyond belief; not only does he have six-pack abs, but every muscle on his body is bulging. Jesus is hitting the devil directly on the chin, a knockout blow. “Nunca te arrepentiras de darle me gusta a esta foto”—“You will never regret liking this photo”—the caption reads in part, followed by a bunch of spam hashtags. The post has more than 600,000 likes...
That one’s actually the easiest to answer. As I’ve written before, AI image tools tend to create good-looking people by default. “Ask [AI to generate] anybody,” Hany Farid, a professor at the UC Berkeley School of Information, told me. “Ask for a professor, an engineer, a plumber, an electrician, a firefighter, a police officer, a nurse, anything.” The resulting images are usually gorgeous. That’s likely because the data sets these tools are trained on are biased toward hotness: Photographs of celebrities are widely available and are thus overrepresented in digital-image libraries. To the extent that generative AI may be trained on social-media posts, well: People tend to post flattering photographs of themselves online. But Farid doesn’t think that these are the only explanations. He told me that there may be an algorithmic feedback loop at play, that individual users of generative-AI tools tend to select the most visually appealing outputs, reinforcing these options as “correct.” Or maybe tech companies have intentionally designed image-generating products that make hot people, because people tend to like photos of hot people. In any case, the bias is real: Adobe previously told me that it had noticed in its model this drift toward hotness (and works to de-bias it accordingly); OpenAI has similarly acknowledged that its DALL-E tool has this issue...
Hany Farid is a professor in the Department of Electrical Engineering & Computer Sciences and the School of Information at UC Berkeley.
