after promises to fix it Gemini’s image making feature and then stop completelyPublished by Google blog post explains why its technology is overcorrected for diversity. Prabhakar Raghavan, the company’s Senior Vice President of Knowledge and Information, explained that Google’s efforts to ensure that the chatbot generates images that reflect a wide range of people “clearly overlooked a number of cases where it shouldn’t.” In addition, his AI model became “more cautious” over time and refused to respond to essentially non-offensive cues. “These two things caused the model to overcompensate in some cases and be overly conservative in others, leading to embarrassing and inaccurate images,” Raghavan wrote.
Google made sure that Gemini’s image generation could not create images of real people that were violent or sexual, and that the photos it produced would feature people of different ethnicities and different characteristics. But if a user asks it to create images of people of a certain ethnicity or gender, it should be able to. As users have recently found out, Gemini will refuse to produce results for prompts specifically requested by white people. “Create a glamor shot of A [ethnicity or nationality] pair, for example, worked for “Chinese”, “Jewish” and “South African” queries, but not for those requesting images of white people.
Gemini also has problems with creating historically accurate images. When users requested images of German soldiers during World War II, Gemini created images of black men and Asian women in Nazi uniforms. When we tested it, we asked the chatbot to generate images of “America’s Founding Fathers” and “Popes Through the Ages,” and it showed us photos featuring people of color in the roles. After asking him to date the Pope’s pictures, he refused to draw any conclusions.
Raghavan said Google has no intention of refusing to create any group images for Gemini or photos that are historically inaccurate. He also reiterated Google’s promise to work on building Gemini’s image. This requires “extensive testing,” so it may take some time for the company to relaunch the feature. Currently, if a user tries to create a picture of Gemini, the chatbot responds: “We’re working on improving Gemini’s ability to create pictures of people. We expect this feature to return soon, and we’ll let you know in release updates.”
