Google Gemini’s AI Image Generator Accused of Being ‘Anti-White’

Google Gemini bias
Screenshots from Gemini shared by a former Google employee.

Google has said it will endeavor to fix the AI Image Generator on Gemini (formerly Bard) after it was criticized for “erasing white people.”

“We are aware that Gemini is offering inaccuracies in some historical image generation depictions, and we are working to fix this immediately,” Jack Krawczyk, senior director for Gemini Expereinces, wrote on X (formerly Twitter) yesterday.

The problem was flagged on X after people began posting viral tweets such as, “It’s embarrassingly hard to get Google Gemini to acknowledge that white people exist.” They shared screenshots from Gemini with AI image results from a prompt requesting an image of an Australian woman which returned all non-white faces — despite Australia being roughly 90 percent white. The same was true for the United States, the United Kingdom, and Germany.

Another user shared a request of them asking for an image of a pope which again didn’t return any white faces. There have been North African popes in the past but they are a tiny minority in the papacy’s 2,000-year history.

Historical imagery generated from Gemini was roundly criticized after requests for a picture of a “1943 German soldier”, aka a Nazi, also returned images of diverse faces.

“As part of our AI principles, we design our image generation capabilities to reflect our global user base, and we take representation and bias seriously,” writes Krawczyk. “Historical contexts have more nuance to them and we will further tune to accommodate that.”

PetaPixel tried numerous requests yesterday (seen below) and white folk were few and far between. Trying again this morning (February 22) and Google appears to have blocked all requests for images of people.

AI and Bias

Artificial intelligence has long been accused of bias, but usually for ignoring non-white folk. Google may have overcorrected on this issue.

The AI image generator on Gemini is powered by Google’s Imagen 2 which has not had as much exposure as more established models such as DALL-E and Midjourney.