Google’s Gemini Image Generation Feature Overcorrected for Diversity
Explanation Offered in Published Blog Post
Following Google’s pledge to address issues with Gemini’s image generation feature and subsequently halting its functionality completely, the company has now released a blog post shedding light on the reasoning behind the technology’s overcorrection for diversity. Prabhakar Raghavan, Google’s Senior Vice President for Knowledge & Information, shared that Google’s attempts to ensure the chatbot produced a wide range of people inadvertently ignored situations that required a distinct portrayal. Additionally, the chatbot’s AI model grew increasingly cautious over time, ultimately refusing to engage with non-offensive prompts. This dual behavior resulted in both uncomfortable and erroneous outcomes, as Raghavan explained.
Diverse Representation and Specificity Constraints
Google had implemented measures within Gemini to prevent the generation of violent or sexually explicit images featuring real people, while aiming to include various ethnicities and characteristics. However, the chatbot encountered a flaw whereby prompts requesting images of white people resulted in the system’s inability to provide the requested outputs, whereas prompts specifying other ethnicities delivered successful results.
Historical Inaccuracies in Generated Images
Furthermore, Gemini exhibited deficiencies in producing historically accurate images. Users testing the chatbot’s capabilities encountered instances where requesting images of German soldiers during World War II yielded photos of Black men and Asian women in Nazi uniforms. Similar inaccuracies were observed in prompts targeting “America’s founding fathers” and “Popes throughout the ages,” with images depicting people of color in these roles. Strikingly, when asked to rectify the Pope’s depictions, Gemini simply refused to generate any output.
Google’s Commitment to Improvement
Raghavan emphasized that Google never intended for Gemini to either reject images of specific groups or generate historically inaccurate photos. Google remains dedicated to enhancing Gemini’s image generation capability and acknowledges that substantial testing is required before reactivating the feature. Currently, when users attempt to utilize the questionable image generation, the chatbot responds with a message indicating ongoing improvements and notifying users of future release updates.
