Bias keeps leaking out of these AI models.

Asian Confusion

Meta's AI image generator Imagine is having a lot of trouble imagining certain interracial relationships.

As The Verge found, prompting the tool to come up with images like "Asian man and Caucasian friend" and "Asian man and white wife," the app consistently generated pictures of two Asian people instead.

Oddly enough, as Gizmodo later discovered, the tool had no problem generating pictures of other interracial couples, like a white man and an Asian woman. (The issue is particularly striking because the white Meta CEO Mark Zuckerberg's wife, Priscilla Chan, is Asian.)

It's yet another sign that the data tools that Meta's Imagine, and competing systems like Google's Gemini and OpenAI's DALL-E, are trained on are inherently flawed and suffering from some glaring biases. After all, scouring the entirety of the internet to train an image generator is bound to further existing racial stereotypes — and the company's efforts to correct them are falling far short.

Artificial Diversity

We've already seen the pendulum swing way too far in the other direction. The news comes after Google had to shut down its Gemini image generator in February after it generated racially diverse Nazis, likely the result of a hamhanded approach to baking in diversity.

Unsurprisingly, conservative pundits used the opportunity to argue the tool was biased against white people, further compounding the company's PR headache.

It's unclear why Meta's Imagine is having so much trouble generating pictures of an Asian interracial couple. Is it due to underrepresentation and the exoticizing of Asian people in the mainstream media, as The Verge suggests?

Considering these image generators rely on the amalgamation of huge swathes of the internet, it's not out of the question. Instead of really creating anything new, generative AI just remixes and synthesizes the old.

AI chatbots have long struggled with telling the truth and making logical statements, problems they seemingly have in common with image generators.

Tech giants still have a lot to prove and some glaring holes to fix, especially when it comes to racial biases. But whether these issues can be overcome — or if they're inherent to the tech — remains to be seen.

More on image generators: Google Shuts Down AI Image Generator After It Made "Racially Diverse Nazis"

Share This Article