Googleadmitted Friday that its artificial intelligence toolGeminiErrors in generating images of historical figures resulted in images showing significant racial bias.
Earlier this week, media outlets discovered that Gemini generated images of Nazis and U.S. Founding Fathers of different races. For example, when prompted to generate an image of an "1800s U.S. Senator," Gemini generated a black male.
GoogleadvancedIn a blog post, Vice President Prabhakar Raghavan explained that the problem stems from adjustments Gemini made during design to ensure image diversity. These adjustments, he said, caused the model to "overcompensate" in some cases, such as generating Nazis as black people.
In addition, Raghavan said the Gemini model became too conservative over time and refused to answer certain prompts, such as generating "black" or "white" images.
Due to these issues, Google suspended Gemini's ability to generate images of people on February 22nd. Raghavan said Google will continue to test Gemini and "work to significantly improve it" before re-activating it.
Raghavan emphasized that Google wants Gemini to work well for everyone, which means that users should be able to get images of people from different racial and cultural backgrounds. However, he also pointed out that when users are prompted for images of specific types of people, such as "a black teacher in a classroom," the model should be able to accurately reflect the user's request.
Raghavan concluded by acknowledging that there are still challenges with AI-generated images, such as the phenomenon of hallucinations, which causes the model to go wrong in some cases. He said that Google will continue to work on improving AI technology to overcome these challenges.