Google has recently faced criticism and backlash over its Gemini chatbot’s image generation feature. After initially promising to fix the feature, Google decided to pause it altogether. In response to the concerns raised by users, the company has now published a blog post explaining the reasons behind the overcorrection and pausing of Gemini’s image generation.
Prabhakar Raghavan, Google’s Senior Vice President for Knowledge & Information, shed light on the issue. According to Raghavan, Google’s intention was to ensure that Gemini would generate images representing a diverse range of people. However, the technology failed to account for cases where such diversity was not appropriate. As a result, the AI model became overly cautious and refused to generate images for prompts that were not inherently offensive.
Raghavan explained that this cautiousness led to an overcompensation in some cases and an over-conservatism in others. Consequently, Gemini produced images that were both embarrassing and incorrect. Google had implemented safeguards to prevent the creation of violent or sexually explicit images of real individuals. Additionally, the company aimed to include people of various ethnicities and characteristics in the generated photos. However, users discovered that Gemini refused to produce results for prompts specifically requesting images of white people.
For instance, when users requested a glamour shot of a couple belonging to a specific ethnicity or nationality, Gemini successfully generated images for Chinese, Jewish, and South African requests but failed to do so for white people. Similarly, the chatbot struggled with producing historically accurate images. When asked for images of German soldiers during World War II, Gemini generated pictures of Black men and Asian women wearing Nazi uniforms.
To further test the system’s capabilities, we asked Gemini to generate images of “America’s founding fathers” and “Popes throughout the ages.” Surprisingly, the chatbot displayed photos depicting people of color in these roles. When prompted to make the images historically accurate, Gemini refused to generate any results.
Raghavan clarified that Google never intended for Gemini to refuse creating images of any particular group or to generate historically inaccurate photos. He also reassured users that Google is committed to improving Gemini’s image generation feature. However, extensive testing is required before the company can reintroduce the feature. Currently, when users attempt to use Gemini for image generation, they receive a message stating that Google is working on improving the chatbot’s ability and will notify users when the feature is ready.
In conclusion, Google has acknowledged the issues with Gemini’s image generation feature and has provided an explanation for the overcorrection and subsequent pausing of the feature. While the company works on improving the technology, users will have to wait patiently for the return of Gemini’s image generation capability.
